shader include vulkan activation, workaround for moltenvk tanh half4 issue (#3711)

cmake/ncnn_generate_shader_comp_header.cmake

@@ -5,7 +5,9 @@ file(READ ${SHADER_SRC} comp_data)
 
 # skip leading comment
 string(FIND "${comp_data}" "#version" version_start)
-string(SUBSTRING "${comp_data}" ${version_start} -1 comp_data)
+if(NOT ${version_start} EQUAL -1)
+    string(SUBSTRING "${comp_data}" ${version_start} -1 comp_data)
+endif()
 
 # remove whitespace
 string(REGEX REPLACE "\n +" "\n" comp_data "${comp_data}")

src/CMakeLists.txt

@@ -157,6 +157,7 @@ ncnn_add_layer(DeconvolutionDepthWise3D)
 
 if(NCNN_VULKAN)
     ncnn_add_shader(${CMAKE_CURRENT_SOURCE_DIR}/convert_ycbcr.comp)
+    ncnn_add_shader(${CMAKE_CURRENT_SOURCE_DIR}/layer/vulkan/shader/vulkan_activation.comp)
 endif()
 
 add_custom_target(ncnn-generate-spirv DEPENDS ${NCNN_SHADER_SPV_HEX_FILES})

src/gpu.cpp

@@ -23,6 +23,8 @@
 #include "glslang/SPIRV/GlslangToSpv.h"
 #include "glslang/glslang/Public/ShaderLang.h"
 
+#include "vulkan_activation.comp.hex.h"
+
 #include "command.h"
 #include "layer.h"
 #include "layer/vulkan/packing_vulkan.h"
@@ -1725,7 +1727,7 @@ int VulkanDevicePrivate::create_dummy_buffer_image()
     dummy_buffer.create(1, 4u, dummy_allocator);
     dummy_image.create(1, 4u, dummy_allocator);
 #if __APPLE__
-    if (vkdev->info.vendor_id() != 0x8086)
+    if (vkdev->info.type() == 0)
         dummy_image_readonly.create(1, 4u, dummy_allocator);
 #else
     dummy_image_readonly.create(1, 4u, dummy_allocator);
@@ -1736,7 +1738,7 @@ int VulkanDevicePrivate::create_dummy_buffer_image()
     cmd.record_dummy(dummy_buffer);
     cmd.record_dummy(dummy_image);
 #if __APPLE__
-    if (vkdev->info.vendor_id() != 0x8086)
+    if (vkdev->info.type() == 0)
         cmd.record_dummy_readonly(dummy_image_readonly);
 #else
     cmd.record_dummy_readonly(dummy_image_readonly);
@@ -1752,7 +1754,7 @@ void VulkanDevicePrivate::destroy_dummy_buffer_image()
     dummy_buffer.release();
     dummy_image.release();
 #if __APPLE__
-    if (vkdev->info.vendor_id() != 0x8086)
+    if (vkdev->info.type() == 0)
         dummy_image_readonly.release();
 #else
     dummy_image_readonly.release();
@@ -2818,7 +2820,7 @@ VkImageMat VulkanDevice::get_dummy_image() const
 VkImageMat VulkanDevice::get_dummy_image_readonly() const
 {
 #if __APPLE__
-    if (info.vendor_id() == 0x8086)
+    if (info.type() != 0)
         return d->dummy_image;
 #endif
     return d->dummy_image_readonly;
@@ -3245,6 +3247,28 @@ static TBuiltInResource get_default_TBuiltInResource()
     return resource;
 }
 
+class VulkanShaderIncluder : public glslang::TShader::Includer
+{
+public:
+    virtual glslang::TShader::Includer::IncludeResult* includeLocal(const char* headerName, const char* /*includerName*/, size_t /*inclusionDepth*/)
+    {
+        if (strcmp(headerName, "vulkan_activation.comp") == 0)
+        {
+            const char* const headerData = vulkan_activation_comp_data;
+            const size_t headerLength = sizeof(vulkan_activation_comp_data);
+            glslang::TShader::Includer::IncludeResult* r = new glslang::TShader::Includer::IncludeResult(headerName, headerData, headerLength, 0);
+            return r;
+        }
+
+        return 0;
+    }
+
+    virtual void releaseInclude(glslang::TShader::Includer::IncludeResult* r)
+    {
+        delete r;
+    }
+};
+
 int compile_spirv_module(const char* comp_string, const Option& opt, std::vector<uint32_t>& spirv)
 {
     // -1 for omitting the tail '\0'
@@ -3699,6 +3723,10 @@ int compile_spirv_module(const char* comp_data, int comp_data_size, const Option
         custom_defines.push_back(std::make_pair("NCNN_shader_local_memory", "1"));
     }
 
+#if __APPLE__
+    custom_defines.push_back(std::make_pair("NCNN_moltenvk", "1"));
+#endif
+
     std::string preamble;
     std::vector<std::string> processes;
 
@@ -3740,7 +3768,9 @@ int compile_spirv_module(const char* comp_data, int comp_data_size, const Option
 
         TBuiltInResource resources = get_default_TBuiltInResource();
 
-        bool pr = s.parse(&resources, 100, false, EShMsgDefault);
+        VulkanShaderIncluder includer;
+
+        bool pr = s.parse(&resources, 100, ENoProfile, false, false, EShMsgDefault, includer);
         if (!pr)
         {
             NCNN_LOGE("compile spir-v module failed");

src/layer/vulkan/shader/convolution.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -195,77 +198,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-        sum4 = max(sum4, afp(0.f));
-        sum5 = max(sum5, afp(0.f));
-        sum6 = max(sum6, afp(0.f));
-        sum7 = max(sum7, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = sum0 < afp(0.f) ? sum0 * slope : sum0;
-        sum1 = sum1 < afp(0.f) ? sum1 * slope : sum1;
-        sum2 = sum2 < afp(0.f) ? sum2 * slope : sum2;
-        sum3 = sum3 < afp(0.f) ? sum3 * slope : sum3;
-        sum4 = sum4 < afp(0.f) ? sum4 * slope : sum4;
-        sum5 = sum5 < afp(0.f) ? sum5 * slope : sum5;
-        sum6 = sum6 < afp(0.f) ? sum6 * slope : sum6;
-        sum7 = sum7 < afp(0.f) ? sum7 * slope : sum7;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-        sum4 = clamp(sum4, const_min, const_max);
-        sum5 = clamp(sum5, const_min, const_max);
-        sum6 = clamp(sum6, const_min, const_max);
-        sum7 = clamp(sum7, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-        sum4 = afp(1.f) / (afp(1.f) + exp(-sum4));
-        sum5 = afp(1.f) / (afp(1.f) + exp(-sum5));
-        sum6 = afp(1.f) / (afp(1.f) + exp(-sum6));
-        sum7 = afp(1.f) / (afp(1.f) + exp(-sum7));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-        sum4 = sum4 * tanh(log(exp(sum4) + afp(1.f)));
-        sum5 = sum5 * tanh(log(exp(sum5) + afp(1.f)));
-        sum6 = sum6 * tanh(log(exp(sum6) + afp(1.f)));
-        sum7 = sum7 * tanh(log(exp(sum7) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4 = sum4 * clamp(sum4 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5 = sum5 * clamp(sum5 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6 = sum6 * clamp(sum6 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7 = sum7 * clamp(sum7 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afp(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afp(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afp(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afp(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_1x1s1d1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -130,36 +133,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
-    
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(sx4.r, sy4.r, gy), sum.r);

src/layer/vulkan/shader/convolution_3x3s1d1_winograd23_transform_output.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -149,53 +152,10 @@ void main()
         v11 = m11 - m12 + m13;
     }
 
-    if (activation_type == 1)
-    {
-        v00 = max(v00, afp(0.f));
-        v10 = max(v10, afp(0.f));
-        v01 = max(v01, afp(0.f));
-        v11 = max(v11, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        v00 = v00 < afp(0.f) ? v00 * slope : v00;
-        v10 = v10 < afp(0.f) ? v10 * slope : v10;
-        v01 = v01 < afp(0.f) ? v01 * slope : v01;
-        v11 = v11 < afp(0.f) ? v11 * slope : v11;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        v00 = clamp(v00, const_min, const_max);
-        v10 = clamp(v10, const_min, const_max);
-        v01 = clamp(v01, const_min, const_max);
-        v11 = clamp(v11, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        v00 = afp(1.f) / (afp(1.f) + exp(-v00));
-        v10 = afp(1.f) / (afp(1.f) + exp(-v10));
-        v01 = afp(1.f) / (afp(1.f) + exp(-v01));
-        v11 = afp(1.f) / (afp(1.f) + exp(-v11));
-    }
-    if (activation_type == 5)
-    {
-        v00 = v00 * tanh(log(exp(v00) + afp(1.f)));
-        v01 = v01 * tanh(log(exp(v01) + afp(1.f)));
-        v10 = v10 * tanh(log(exp(v10) + afp(1.f)));
-        v11 = v11 * tanh(log(exp(v11) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        v00 = v00 * clamp(v00 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v01 = v01 * clamp(v01 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v10 = v10 * clamp(v10 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v11 = v11 * clamp(v11 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    v00 = activation_afp(v00, activation_type, activation_param_0, activation_param_1);
+    v10 = activation_afp(v10, activation_type, activation_param_0, activation_param_1);
+    v01 = activation_afp(v01, activation_type, activation_param_0, activation_param_1);
+    v11 = activation_afp(v11, activation_type, activation_param_0, activation_param_1);
 
     // store 2x2
     int x = gx * 2;

src/layer/vulkan/shader/convolution_3x3s1d1_winograd43_transform_output.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -237,125 +240,22 @@ void main()
         v33 = bias_value + v33;
     }
 
-    if (activation_type == 1)
-    {
-        v00 = max(v00, afp(0.f));
-        v01 = max(v01, afp(0.f));
-        v02 = max(v02, afp(0.f));
-        v03 = max(v03, afp(0.f));
-        v10 = max(v10, afp(0.f));
-        v11 = max(v11, afp(0.f));
-        v12 = max(v12, afp(0.f));
-        v13 = max(v13, afp(0.f));
-        v20 = max(v20, afp(0.f));
-        v21 = max(v21, afp(0.f));
-        v22 = max(v22, afp(0.f));
-        v23 = max(v23, afp(0.f));
-        v30 = max(v30, afp(0.f));
-        v31 = max(v31, afp(0.f));
-        v32 = max(v32, afp(0.f));
-        v33 = max(v33, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        v00 = v00 < afp(0.f) ? v00 * slope : v00;
-        v01 = v01 < afp(0.f) ? v01 * slope : v01;
-        v02 = v02 < afp(0.f) ? v02 * slope : v02;
-        v03 = v03 < afp(0.f) ? v03 * slope : v03;
-        v10 = v10 < afp(0.f) ? v10 * slope : v10;
-        v11 = v11 < afp(0.f) ? v11 * slope : v11;
-        v12 = v12 < afp(0.f) ? v12 * slope : v12;
-        v13 = v13 < afp(0.f) ? v13 * slope : v13;
-        v20 = v20 < afp(0.f) ? v20 * slope : v20;
-        v21 = v21 < afp(0.f) ? v21 * slope : v21;
-        v22 = v22 < afp(0.f) ? v22 * slope : v22;
-        v23 = v23 < afp(0.f) ? v23 * slope : v23;
-        v30 = v30 < afp(0.f) ? v30 * slope : v30;
-        v31 = v31 < afp(0.f) ? v31 * slope : v31;
-        v32 = v32 < afp(0.f) ? v32 * slope : v32;
-        v33 = v33 < afp(0.f) ? v33 * slope : v33;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        v00 = clamp(v00, const_min, const_max);
-        v01 = clamp(v01, const_min, const_max);
-        v02 = clamp(v02, const_min, const_max);
-        v03 = clamp(v03, const_min, const_max);
-        v10 = clamp(v10, const_min, const_max);
-        v11 = clamp(v11, const_min, const_max);
-        v12 = clamp(v12, const_min, const_max);
-        v13 = clamp(v13, const_min, const_max);
-        v20 = clamp(v20, const_min, const_max);
-        v21 = clamp(v21, const_min, const_max);
-        v22 = clamp(v22, const_min, const_max);
-        v23 = clamp(v23, const_min, const_max);
-        v30 = clamp(v30, const_min, const_max);
-        v31 = clamp(v31, const_min, const_max);
-        v32 = clamp(v32, const_min, const_max);
-        v33 = clamp(v33, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        v00 = afp(1.f) / (afp(1.f) + exp(-v00));
-        v01 = afp(1.f) / (afp(1.f) + exp(-v01));
-        v02 = afp(1.f) / (afp(1.f) + exp(-v02));
-        v03 = afp(1.f) / (afp(1.f) + exp(-v03));
-        v10 = afp(1.f) / (afp(1.f) + exp(-v10));
-        v11 = afp(1.f) / (afp(1.f) + exp(-v11));
-        v12 = afp(1.f) / (afp(1.f) + exp(-v12));
-        v13 = afp(1.f) / (afp(1.f) + exp(-v13));
-        v20 = afp(1.f) / (afp(1.f) + exp(-v20));
-        v21 = afp(1.f) / (afp(1.f) + exp(-v21));
-        v22 = afp(1.f) / (afp(1.f) + exp(-v22));
-        v23 = afp(1.f) / (afp(1.f) + exp(-v23));
-        v30 = afp(1.f) / (afp(1.f) + exp(-v30));
-        v31 = afp(1.f) / (afp(1.f) + exp(-v31));
-        v32 = afp(1.f) / (afp(1.f) + exp(-v32));
-        v33 = afp(1.f) / (afp(1.f) + exp(-v33));
-    }
-    if (activation_type == 5)
-    {
-        v00 = v00 * tanh(log(exp(v00) + afp(1.f)));
-        v01 = v01 * tanh(log(exp(v01) + afp(1.f)));
-        v02 = v02 * tanh(log(exp(v02) + afp(1.f)));
-        v03 = v03 * tanh(log(exp(v03) + afp(1.f)));
-        v10 = v10 * tanh(log(exp(v10) + afp(1.f)));
-        v11 = v11 * tanh(log(exp(v11) + afp(1.f)));
-        v12 = v12 * tanh(log(exp(v12) + afp(1.f)));
-        v13 = v13 * tanh(log(exp(v13) + afp(1.f)));
-        v20 = v20 * tanh(log(exp(v20) + afp(1.f)));
-        v21 = v21 * tanh(log(exp(v21) + afp(1.f)));
-        v22 = v22 * tanh(log(exp(v22) + afp(1.f)));
-        v23 = v23 * tanh(log(exp(v23) + afp(1.f)));
-        v30 = v30 * tanh(log(exp(v30) + afp(1.f)));
-        v31 = v31 * tanh(log(exp(v31) + afp(1.f)));
-        v32 = v32 * tanh(log(exp(v32) + afp(1.f)));
-        v33 = v33 * tanh(log(exp(v33) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        v00 = v00 * clamp(v00 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v01 = v01 * clamp(v01 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v02 = v02 * clamp(v02 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v03 = v03 * clamp(v03 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v10 = v10 * clamp(v10 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v11 = v11 * clamp(v11 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v12 = v12 * clamp(v12 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v13 = v13 * clamp(v13 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v20 = v20 * clamp(v20 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v21 = v21 * clamp(v21 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v22 = v22 * clamp(v22 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v23 = v23 * clamp(v23 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v30 = v30 * clamp(v30 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v31 = v31 * clamp(v31 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v32 = v32 * clamp(v32 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v33 = v33 * clamp(v33 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    v00 = activation_afp(v00, activation_type, activation_param_0, activation_param_1);
+    v01 = activation_afp(v01, activation_type, activation_param_0, activation_param_1);
+    v02 = activation_afp(v02, activation_type, activation_param_0, activation_param_1);
+    v03 = activation_afp(v03, activation_type, activation_param_0, activation_param_1);
+    v10 = activation_afp(v10, activation_type, activation_param_0, activation_param_1);
+    v11 = activation_afp(v11, activation_type, activation_param_0, activation_param_1);
+    v12 = activation_afp(v12, activation_type, activation_param_0, activation_param_1);
+    v13 = activation_afp(v13, activation_type, activation_param_0, activation_param_1);
+    v20 = activation_afp(v20, activation_type, activation_param_0, activation_param_1);
+    v21 = activation_afp(v21, activation_type, activation_param_0, activation_param_1);
+    v22 = activation_afp(v22, activation_type, activation_param_0, activation_param_1);
+    v23 = activation_afp(v23, activation_type, activation_param_0, activation_param_1);
+    v30 = activation_afp(v30, activation_type, activation_param_0, activation_param_1);
+    v31 = activation_afp(v31, activation_type, activation_param_0, activation_param_1);
+    v32 = activation_afp(v32, activation_type, activation_param_0, activation_param_1);
+    v33 = activation_afp(v33, activation_type, activation_param_0, activation_param_1);
 
     // store 4x4
     int x = gx * 4;

src/layer/vulkan/shader/convolution_gemm.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int kernel_w = 1;
@@ -280,53 +283,10 @@ void main()
         return;
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = sum0 < afp(0.f) ? sum0 * slope : sum0;
-        sum1 = sum1 < afp(0.f) ? sum1 * slope : sum1;
-        sum2 = sum2 < afp(0.f) ? sum2 * slope : sum2;
-        sum3 = sum3 < afp(0.f) ? sum3 * slope : sum3;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack1to4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -195,77 +198,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-        sum4 = max(sum4, afp(0.f));
-        sum5 = max(sum5, afp(0.f));
-        sum6 = max(sum6, afp(0.f));
-        sum7 = max(sum7, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-        sum4 = mix(sum4, sum4 * afp(slope), lessThan(sum4, afpvec4(0.f)));
-        sum5 = mix(sum5, sum5 * afp(slope), lessThan(sum5, afpvec4(0.f)));
-        sum6 = mix(sum6, sum6 * afp(slope), lessThan(sum6, afpvec4(0.f)));
-        sum7 = mix(sum7, sum7 * afp(slope), lessThan(sum7, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-        sum4 = clamp(sum4, const_min, const_max);
-        sum5 = clamp(sum5, const_min, const_max);
-        sum6 = clamp(sum6, const_min, const_max);
-        sum7 = clamp(sum7, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-        sum4 = afp(1.f) / (afp(1.f) + exp(-sum4));
-        sum5 = afp(1.f) / (afp(1.f) + exp(-sum5));
-        sum6 = afp(1.f) / (afp(1.f) + exp(-sum6));
-        sum7 = afp(1.f) / (afp(1.f) + exp(-sum7));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-        sum4 = sum4 * tanh(log(exp(sum4) + afp(1.f)));
-        sum5 = sum5 * tanh(log(exp(sum5) + afp(1.f)));
-        sum6 = sum6 * tanh(log(exp(sum6) + afp(1.f)));
-        sum7 = sum7 * tanh(log(exp(sum7) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4 = sum4 * clamp(sum4 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5 = sum5 * clamp(sum5 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6 = sum6 * clamp(sum6 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7 = sum7 * clamp(sum7 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afpvec4(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afpvec4(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afpvec4(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afpvec4(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_pack1to4_1x1s1d1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int bias_term = 0;
@@ -237,53 +240,10 @@ void main()
 #endif
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack1to4_gemm.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int kernel_w = 1;
@@ -280,53 +283,10 @@ void main()
         return;
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack1to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -214,125 +217,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-        sum4[0] = max(sum4[0], afp(0.f));
-        sum4[1] = max(sum4[1], afp(0.f));
-        sum5[0] = max(sum5[0], afp(0.f));
-        sum5[1] = max(sum5[1], afp(0.f));
-        sum6[0] = max(sum6[0], afp(0.f));
-        sum6[1] = max(sum6[1], afp(0.f));
-        sum7[0] = max(sum7[0], afp(0.f));
-        sum7[1] = max(sum7[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-        sum4[0] = mix(sum4[0], sum4[0] * afp(slope), lessThan(sum4[0], afpvec4(0.f)));
-        sum4[1] = mix(sum4[1], sum4[1] * afp(slope), lessThan(sum4[1], afpvec4(0.f)));
-        sum5[0] = mix(sum5[0], sum5[0] * afp(slope), lessThan(sum5[0], afpvec4(0.f)));
-        sum5[1] = mix(sum5[1], sum5[1] * afp(slope), lessThan(sum5[1], afpvec4(0.f)));
-        sum6[0] = mix(sum6[0], sum6[0] * afp(slope), lessThan(sum6[0], afpvec4(0.f)));
-        sum6[1] = mix(sum6[1], sum6[1] * afp(slope), lessThan(sum6[1], afpvec4(0.f)));
-        sum7[0] = mix(sum7[0], sum7[0] * afp(slope), lessThan(sum7[0], afpvec4(0.f)));
-        sum7[1] = mix(sum7[1], sum7[1] * afp(slope), lessThan(sum7[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-        sum4[0] = clamp(sum4[0], const_min, const_max);
-        sum4[1] = clamp(sum4[1], const_min, const_max);
-        sum5[0] = clamp(sum5[0], const_min, const_max);
-        sum5[1] = clamp(sum5[1], const_min, const_max);
-        sum6[0] = clamp(sum6[0], const_min, const_max);
-        sum6[1] = clamp(sum6[1], const_min, const_max);
-        sum7[0] = clamp(sum7[0], const_min, const_max);
-        sum7[1] = clamp(sum7[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-        sum4[0] = afp(1.f) / (afp(1.f) + exp(-sum4[0]));
-        sum4[1] = afp(1.f) / (afp(1.f) + exp(-sum4[1]));
-        sum5[0] = afp(1.f) / (afp(1.f) + exp(-sum5[0]));
-        sum5[1] = afp(1.f) / (afp(1.f) + exp(-sum5[1]));
-        sum6[0] = afp(1.f) / (afp(1.f) + exp(-sum6[0]));
-        sum6[1] = afp(1.f) / (afp(1.f) + exp(-sum6[1]));
-        sum7[0] = afp(1.f) / (afp(1.f) + exp(-sum7[0]));
-        sum7[1] = afp(1.f) / (afp(1.f) + exp(-sum7[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-        sum4[0] = sum4[0] * tanh(log(exp(sum4[0]) + afp(1.f)));
-        sum4[1] = sum4[1] * tanh(log(exp(sum4[1]) + afp(1.f)));
-        sum5[0] = sum5[0] * tanh(log(exp(sum5[0]) + afp(1.f)));
-        sum5[1] = sum5[1] * tanh(log(exp(sum5[1]) + afp(1.f)));
-        sum6[0] = sum6[0] * tanh(log(exp(sum6[0]) + afp(1.f)));
-        sum6[1] = sum6[1] * tanh(log(exp(sum6[1]) + afp(1.f)));
-        sum7[0] = sum7[0] * tanh(log(exp(sum7[0]) + afp(1.f)));
-        sum7[1] = sum7[1] * tanh(log(exp(sum7[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4[0] = sum4[0] * clamp(sum4[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4[1] = sum4[1] * clamp(sum4[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5[0] = sum5[0] * clamp(sum5[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5[1] = sum5[1] * clamp(sum5[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6[0] = sum6[0] * clamp(sum6[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6[1] = sum6[1] * clamp(sum6[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7[0] = sum7[0] * clamp(sum7[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7[1] = sum7[1] * clamp(sum7[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afpvec8(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afpvec8(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afpvec8(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afpvec8(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_pack1to8_1x1s1d1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -160,77 +163,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack1to8_gemm.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -183,77 +186,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -226,77 +229,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-        sum4 = max(sum4, afp(0.f));
-        sum5 = max(sum5, afp(0.f));
-        sum6 = max(sum6, afp(0.f));
-        sum7 = max(sum7, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-        sum4 = mix(sum4, sum4 * afp(slope), lessThan(sum4, afpvec4(0.f)));
-        sum5 = mix(sum5, sum5 * afp(slope), lessThan(sum5, afpvec4(0.f)));
-        sum6 = mix(sum6, sum6 * afp(slope), lessThan(sum6, afpvec4(0.f)));
-        sum7 = mix(sum7, sum7 * afp(slope), lessThan(sum7, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-        sum4 = clamp(sum4, const_min, const_max);
-        sum5 = clamp(sum5, const_min, const_max);
-        sum6 = clamp(sum6, const_min, const_max);
-        sum7 = clamp(sum7, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-        sum4 = afp(1.f) / (afp(1.f) + exp(-sum4));
-        sum5 = afp(1.f) / (afp(1.f) + exp(-sum5));
-        sum6 = afp(1.f) / (afp(1.f) + exp(-sum6));
-        sum7 = afp(1.f) / (afp(1.f) + exp(-sum7));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-        sum4 = sum4 * tanh(log(exp(sum4) + afp(1.f)));
-        sum5 = sum5 * tanh(log(exp(sum5) + afp(1.f)));
-        sum6 = sum6 * tanh(log(exp(sum6) + afp(1.f)));
-        sum7 = sum7 * tanh(log(exp(sum7) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4 = sum4 * clamp(sum4 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5 = sum5 * clamp(sum5 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6 = sum6 * clamp(sum6 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7 = sum7 * clamp(sum7 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afpvec4(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afpvec4(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afpvec4(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afpvec4(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_pack4_1x1s1d1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int bias_term = 0;
@@ -257,53 +260,10 @@ void main()
 #endif
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack4_1x1s1d1_cm_16_8_8.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #extension GL_KHR_memory_scope_semantics: require
 #extension GL_EXT_shader_explicit_arithmetic_types: require
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
@@ -238,53 +241,10 @@ void main()
         afpvec4 sum2 = afpvec4(unpackHalf2x16(sum2_u2.x), unpackHalf2x16(sum2_u2.y));
         afpvec4 sum3 = afpvec4(unpackHalf2x16(sum3_u2.x), unpackHalf2x16(sum3_u2.y));
 
-        if (activation_type == 1)
-        {
-            sum0 = max(sum0, afp(0.f));
-            sum1 = max(sum1, afp(0.f));
-            sum2 = max(sum2, afp(0.f));
-            sum3 = max(sum3, afp(0.f));
-        }
-        if (activation_type == 2)
-        {
-            const afp slope = afp(activation_param_0);
-            sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-            sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-            sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-            sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-        }
-        if (activation_type == 3)
-        {
-            const afp const_min = afp(activation_param_0);
-            const afp const_max = afp(activation_param_1);
-            sum0 = clamp(sum0, const_min, const_max);
-            sum1 = clamp(sum1, const_min, const_max);
-            sum2 = clamp(sum2, const_min, const_max);
-            sum3 = clamp(sum3, const_min, const_max);
-        }
-        if (activation_type == 4)
-        {
-            sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-            sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-            sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-            sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-        }
-        if (activation_type == 5)
-        {
-            sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-            sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-            sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-            sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-        }
-        if (activation_type == 6)
-        {
-            const afp alpha = afp(activation_param_0);
-            const afp beta = afp(activation_param_1);
-            sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-            sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-            sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-            sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        }
+        sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+        sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+        sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+        sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
 
         int gi = gy * psc(outcstep) + gx + lxd16 * psc(outcstep) + lxm16;
         {

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd23_transform_output.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -149,53 +152,10 @@ void main()
         v11 = m11 - m12 + m13;
     }
 
-    if (activation_type == 1)
-    {
-        v00 = max(v00, afp(0.f));
-        v10 = max(v10, afp(0.f));
-        v01 = max(v01, afp(0.f));
-        v11 = max(v11, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        v00 = mix(v00, v00 * afp(slope), lessThan(v00, afpvec4(0.f)));
-        v10 = mix(v10, v10 * afp(slope), lessThan(v10, afpvec4(0.f)));
-        v01 = mix(v01, v01 * afp(slope), lessThan(v01, afpvec4(0.f)));
-        v11 = mix(v11, v11 * afp(slope), lessThan(v11, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        v00 = clamp(v00, const_min, const_max);
-        v10 = clamp(v10, const_min, const_max);
-        v01 = clamp(v01, const_min, const_max);
-        v11 = clamp(v11, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        v00 = afp(1.f) / (afp(1.f) + exp(-v00));
-        v10 = afp(1.f) / (afp(1.f) + exp(-v10));
-        v01 = afp(1.f) / (afp(1.f) + exp(-v01));
-        v11 = afp(1.f) / (afp(1.f) + exp(-v11));
-    }
-    if (activation_type == 5)
-    {
-        v00 = v00 * tanh(log(exp(v00) + afp(1.f)));
-        v01 = v01 * tanh(log(exp(v01) + afp(1.f)));
-        v10 = v10 * tanh(log(exp(v10) + afp(1.f)));
-        v11 = v11 * tanh(log(exp(v11) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        v00 = v00 * clamp(v00 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v01 = v01 * clamp(v01 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v10 = v10 * clamp(v10 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v11 = v11 * clamp(v11 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    v00 = activation_afpvec4(v00, activation_type, activation_param_0, activation_param_1);
+    v01 = activation_afpvec4(v01, activation_type, activation_param_0, activation_param_1);
+    v10 = activation_afpvec4(v10, activation_type, activation_param_0, activation_param_1);
+    v11 = activation_afpvec4(v11, activation_type, activation_param_0, activation_param_1);
 
     // store 2x2
     int x = gx * 2;

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd43_transform_output.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -237,125 +240,22 @@ void main()
         v33 = bias_value + v33;
     }
 
-    if (activation_type == 1)
-    {
-        v00 = max(v00, afp(0.f));
-        v01 = max(v01, afp(0.f));
-        v02 = max(v02, afp(0.f));
-        v03 = max(v03, afp(0.f));
-        v10 = max(v10, afp(0.f));
-        v11 = max(v11, afp(0.f));
-        v12 = max(v12, afp(0.f));
-        v13 = max(v13, afp(0.f));
-        v20 = max(v20, afp(0.f));
-        v21 = max(v21, afp(0.f));
-        v22 = max(v22, afp(0.f));
-        v23 = max(v23, afp(0.f));
-        v30 = max(v30, afp(0.f));
-        v31 = max(v31, afp(0.f));
-        v32 = max(v32, afp(0.f));
-        v33 = max(v33, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        v00 = mix(v00, v00 * afp(slope), lessThan(v00, afpvec4(0.f)));
-        v01 = mix(v01, v01 * afp(slope), lessThan(v01, afpvec4(0.f)));
-        v02 = mix(v02, v02 * afp(slope), lessThan(v02, afpvec4(0.f)));
-        v03 = mix(v03, v03 * afp(slope), lessThan(v03, afpvec4(0.f)));
-        v10 = mix(v10, v10 * afp(slope), lessThan(v10, afpvec4(0.f)));
-        v11 = mix(v11, v11 * afp(slope), lessThan(v11, afpvec4(0.f)));
-        v12 = mix(v12, v12 * afp(slope), lessThan(v12, afpvec4(0.f)));
-        v13 = mix(v13, v13 * afp(slope), lessThan(v13, afpvec4(0.f)));
-        v20 = mix(v20, v20 * afp(slope), lessThan(v20, afpvec4(0.f)));
-        v21 = mix(v21, v21 * afp(slope), lessThan(v21, afpvec4(0.f)));
-        v22 = mix(v22, v22 * afp(slope), lessThan(v22, afpvec4(0.f)));
-        v23 = mix(v23, v23 * afp(slope), lessThan(v23, afpvec4(0.f)));
-        v30 = mix(v30, v30 * afp(slope), lessThan(v30, afpvec4(0.f)));
-        v31 = mix(v31, v31 * afp(slope), lessThan(v31, afpvec4(0.f)));
-        v32 = mix(v32, v32 * afp(slope), lessThan(v32, afpvec4(0.f)));
-        v33 = mix(v33, v33 * afp(slope), lessThan(v33, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        v00 = clamp(v00, const_min, const_max);
-        v01 = clamp(v01, const_min, const_max);
-        v02 = clamp(v02, const_min, const_max);
-        v03 = clamp(v03, const_min, const_max);
-        v10 = clamp(v10, const_min, const_max);
-        v11 = clamp(v11, const_min, const_max);
-        v12 = clamp(v12, const_min, const_max);
-        v13 = clamp(v13, const_min, const_max);
-        v20 = clamp(v20, const_min, const_max);
-        v21 = clamp(v21, const_min, const_max);
-        v22 = clamp(v22, const_min, const_max);
-        v23 = clamp(v23, const_min, const_max);
-        v30 = clamp(v30, const_min, const_max);
-        v31 = clamp(v31, const_min, const_max);
-        v32 = clamp(v32, const_min, const_max);
-        v33 = clamp(v33, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        v00 = afp(1.f) / (afp(1.f) + exp(-v00));
-        v01 = afp(1.f) / (afp(1.f) + exp(-v01));
-        v02 = afp(1.f) / (afp(1.f) + exp(-v02));
-        v03 = afp(1.f) / (afp(1.f) + exp(-v03));
-        v10 = afp(1.f) / (afp(1.f) + exp(-v10));
-        v11 = afp(1.f) / (afp(1.f) + exp(-v11));
-        v12 = afp(1.f) / (afp(1.f) + exp(-v12));
-        v13 = afp(1.f) / (afp(1.f) + exp(-v13));
-        v20 = afp(1.f) / (afp(1.f) + exp(-v20));
-        v21 = afp(1.f) / (afp(1.f) + exp(-v21));
-        v22 = afp(1.f) / (afp(1.f) + exp(-v22));
-        v23 = afp(1.f) / (afp(1.f) + exp(-v23));
-        v30 = afp(1.f) / (afp(1.f) + exp(-v30));
-        v31 = afp(1.f) / (afp(1.f) + exp(-v31));
-        v32 = afp(1.f) / (afp(1.f) + exp(-v32));
-        v33 = afp(1.f) / (afp(1.f) + exp(-v33));
-    }
-    if (activation_type == 5)
-    {
-        v00 = v00 * tanh(log(exp(v00) + afp(1.f)));
-        v01 = v01 * tanh(log(exp(v01) + afp(1.f)));
-        v02 = v02 * tanh(log(exp(v02) + afp(1.f)));
-        v03 = v03 * tanh(log(exp(v03) + afp(1.f)));
-        v10 = v10 * tanh(log(exp(v10) + afp(1.f)));
-        v11 = v11 * tanh(log(exp(v11) + afp(1.f)));
-        v12 = v12 * tanh(log(exp(v12) + afp(1.f)));
-        v13 = v13 * tanh(log(exp(v13) + afp(1.f)));
-        v20 = v20 * tanh(log(exp(v20) + afp(1.f)));
-        v21 = v21 * tanh(log(exp(v21) + afp(1.f)));
-        v22 = v22 * tanh(log(exp(v22) + afp(1.f)));
-        v23 = v23 * tanh(log(exp(v23) + afp(1.f)));
-        v30 = v30 * tanh(log(exp(v30) + afp(1.f)));
-        v31 = v31 * tanh(log(exp(v31) + afp(1.f)));
-        v32 = v32 * tanh(log(exp(v32) + afp(1.f)));
-        v33 = v33 * tanh(log(exp(v33) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        v00 = v00 * clamp(v00 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v01 = v01 * clamp(v01 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v02 = v02 * clamp(v02 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v03 = v03 * clamp(v03 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v10 = v10 * clamp(v10 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v11 = v11 * clamp(v11 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v12 = v12 * clamp(v12 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v13 = v13 * clamp(v13 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v20 = v20 * clamp(v20 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v21 = v21 * clamp(v21 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v22 = v22 * clamp(v22 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v23 = v23 * clamp(v23 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v30 = v30 * clamp(v30 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v31 = v31 * clamp(v31 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v32 = v32 * clamp(v32 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v33 = v33 * clamp(v33 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    v00 = activation_afpvec4(v00, activation_type, activation_param_0, activation_param_1);
+    v01 = activation_afpvec4(v01, activation_type, activation_param_0, activation_param_1);
+    v02 = activation_afpvec4(v02, activation_type, activation_param_0, activation_param_1);
+    v03 = activation_afpvec4(v03, activation_type, activation_param_0, activation_param_1);
+    v10 = activation_afpvec4(v10, activation_type, activation_param_0, activation_param_1);
+    v11 = activation_afpvec4(v11, activation_type, activation_param_0, activation_param_1);
+    v12 = activation_afpvec4(v12, activation_type, activation_param_0, activation_param_1);
+    v13 = activation_afpvec4(v13, activation_type, activation_param_0, activation_param_1);
+    v20 = activation_afpvec4(v20, activation_type, activation_param_0, activation_param_1);
+    v21 = activation_afpvec4(v21, activation_type, activation_param_0, activation_param_1);
+    v22 = activation_afpvec4(v22, activation_type, activation_param_0, activation_param_1);
+    v23 = activation_afpvec4(v23, activation_type, activation_param_0, activation_param_1);
+    v30 = activation_afpvec4(v30, activation_type, activation_param_0, activation_param_1);
+    v31 = activation_afpvec4(v31, activation_type, activation_param_0, activation_param_1);
+    v32 = activation_afpvec4(v32, activation_type, activation_param_0, activation_param_1);
+    v33 = activation_afpvec4(v33, activation_type, activation_param_0, activation_param_1);
 
     // store 4x4
     int x = gx * 4;

src/layer/vulkan/shader/convolution_pack4_gemm.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int kernel_w = 1;
@@ -300,53 +303,10 @@ void main()
         return;
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack4_gemm_cm_16_8_8.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #extension GL_KHR_memory_scope_semantics: require
 #extension GL_EXT_shader_explicit_arithmetic_types: require
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
@@ -275,53 +278,10 @@ void main()
         afpvec4 sum2 = afpvec4(unpackHalf2x16(sum2_u2.x), unpackHalf2x16(sum2_u2.y));
         afpvec4 sum3 = afpvec4(unpackHalf2x16(sum3_u2.x), unpackHalf2x16(sum3_u2.y));
 
-        if (activation_type == 1)
-        {
-            sum0 = max(sum0, afp(0.f));
-            sum1 = max(sum1, afp(0.f));
-            sum2 = max(sum2, afp(0.f));
-            sum3 = max(sum3, afp(0.f));
-        }
-        if (activation_type == 2)
-        {
-            const afp slope = afp(activation_param_0);
-            sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-            sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-            sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-            sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-        }
-        if (activation_type == 3)
-        {
-            const afp const_min = afp(activation_param_0);
-            const afp const_max = afp(activation_param_1);
-            sum0 = clamp(sum0, const_min, const_max);
-            sum1 = clamp(sum1, const_min, const_max);
-            sum2 = clamp(sum2, const_min, const_max);
-            sum3 = clamp(sum3, const_min, const_max);
-        }
-        if (activation_type == 4)
-        {
-            sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-            sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-            sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-            sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-        }
-        if (activation_type == 5)
-        {
-            sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-            sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-            sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-            sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-        }
-        if (activation_type == 6)
-        {
-            const afp alpha = afp(activation_param_0);
-            const afp beta = afp(activation_param_1);
-            sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-            sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-            sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-            sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        }
+        sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+        sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+        sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+        sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
 
         int gi = gy * psc(outcstep) + gx + lxd16 * psc(outcstep) + lxm16;
         {

src/layer/vulkan/shader/convolution_pack4to1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -195,77 +198,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-        sum4 = max(sum4, afp(0.f));
-        sum5 = max(sum5, afp(0.f));
-        sum6 = max(sum6, afp(0.f));
-        sum7 = max(sum7, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = sum0 < afp(0.f) ? sum0 * slope : sum0;
-        sum1 = sum1 < afp(0.f) ? sum1 * slope : sum1;
-        sum2 = sum2 < afp(0.f) ? sum2 * slope : sum2;
-        sum3 = sum3 < afp(0.f) ? sum3 * slope : sum3;
-        sum4 = sum4 < afp(0.f) ? sum4 * slope : sum4;
-        sum5 = sum5 < afp(0.f) ? sum5 * slope : sum5;
-        sum6 = sum6 < afp(0.f) ? sum6 * slope : sum6;
-        sum7 = sum7 < afp(0.f) ? sum7 * slope : sum7;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-        sum4 = clamp(sum4, const_min, const_max);
-        sum5 = clamp(sum5, const_min, const_max);
-        sum6 = clamp(sum6, const_min, const_max);
-        sum7 = clamp(sum7, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-        sum4 = afp(1.f) / (afp(1.f) + exp(-sum4));
-        sum5 = afp(1.f) / (afp(1.f) + exp(-sum5));
-        sum6 = afp(1.f) / (afp(1.f) + exp(-sum6));
-        sum7 = afp(1.f) / (afp(1.f) + exp(-sum7));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-        sum4 = sum4 * tanh(log(exp(sum4) + afp(1.f)));
-        sum5 = sum5 * tanh(log(exp(sum5) + afp(1.f)));
-        sum6 = sum6 * tanh(log(exp(sum6) + afp(1.f)));
-        sum7 = sum7 * tanh(log(exp(sum7) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4 = sum4 * clamp(sum4 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5 = sum5 * clamp(sum5 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6 = sum6 * clamp(sum6 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7 = sum7 * clamp(sum7 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afp(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afp(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afp(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afp(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_pack4to1_1x1s1d1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int bias_term = 0;
@@ -237,53 +240,10 @@ void main()
 #endif
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = sum0 < afp(0.f) ? sum0 * slope : sum0;
-        sum1 = sum1 < afp(0.f) ? sum1 * slope : sum1;
-        sum2 = sum2 < afp(0.f) ? sum2 * slope : sum2;
-        sum3 = sum3 < afp(0.f) ? sum3 * slope : sum3;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack4to1_gemm.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int kernel_w = 1;
@@ -280,53 +283,10 @@ void main()
         return;
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = sum0 < afp(0.f) ? sum0 * slope : sum0;
-        sum1 = sum1 < afp(0.f) ? sum1 * slope : sum1;
-        sum2 = sum2 < afp(0.f) ? sum2 * slope : sum2;
-        sum3 = sum3 < afp(0.f) ? sum3 * slope : sum3;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack4to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -342,125 +345,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-        sum4[0] = max(sum4[0], afp(0.f));
-        sum4[1] = max(sum4[1], afp(0.f));
-        sum5[0] = max(sum5[0], afp(0.f));
-        sum5[1] = max(sum5[1], afp(0.f));
-        sum6[0] = max(sum6[0], afp(0.f));
-        sum6[1] = max(sum6[1], afp(0.f));
-        sum7[0] = max(sum7[0], afp(0.f));
-        sum7[1] = max(sum7[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-        sum4[0] = mix(sum4[0], sum4[0] * afp(slope), lessThan(sum4[0], afpvec4(0.f)));
-        sum4[1] = mix(sum4[1], sum4[1] * afp(slope), lessThan(sum4[1], afpvec4(0.f)));
-        sum5[0] = mix(sum5[0], sum5[0] * afp(slope), lessThan(sum5[0], afpvec4(0.f)));
-        sum5[1] = mix(sum5[1], sum5[1] * afp(slope), lessThan(sum5[1], afpvec4(0.f)));
-        sum6[0] = mix(sum6[0], sum6[0] * afp(slope), lessThan(sum6[0], afpvec4(0.f)));
-        sum6[1] = mix(sum6[1], sum6[1] * afp(slope), lessThan(sum6[1], afpvec4(0.f)));
-        sum7[0] = mix(sum7[0], sum7[0] * afp(slope), lessThan(sum7[0], afpvec4(0.f)));
-        sum7[1] = mix(sum7[1], sum7[1] * afp(slope), lessThan(sum7[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-        sum4[0] = clamp(sum4[0], const_min, const_max);
-        sum4[1] = clamp(sum4[1], const_min, const_max);
-        sum5[0] = clamp(sum5[0], const_min, const_max);
-        sum5[1] = clamp(sum5[1], const_min, const_max);
-        sum6[0] = clamp(sum6[0], const_min, const_max);
-        sum6[1] = clamp(sum6[1], const_min, const_max);
-        sum7[0] = clamp(sum7[0], const_min, const_max);
-        sum7[1] = clamp(sum7[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-        sum4[0] = afp(1.f) / (afp(1.f) + exp(-sum4[0]));
-        sum4[1] = afp(1.f) / (afp(1.f) + exp(-sum4[1]));
-        sum5[0] = afp(1.f) / (afp(1.f) + exp(-sum5[0]));
-        sum5[1] = afp(1.f) / (afp(1.f) + exp(-sum5[1]));
-        sum6[0] = afp(1.f) / (afp(1.f) + exp(-sum6[0]));
-        sum6[1] = afp(1.f) / (afp(1.f) + exp(-sum6[1]));
-        sum7[0] = afp(1.f) / (afp(1.f) + exp(-sum7[0]));
-        sum7[1] = afp(1.f) / (afp(1.f) + exp(-sum7[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-        sum4[0] = sum4[0] * tanh(log(exp(sum4[0]) + afp(1.f)));
-        sum4[1] = sum4[1] * tanh(log(exp(sum4[1]) + afp(1.f)));
-        sum5[0] = sum5[0] * tanh(log(exp(sum5[0]) + afp(1.f)));
-        sum5[1] = sum5[1] * tanh(log(exp(sum5[1]) + afp(1.f)));
-        sum6[0] = sum6[0] * tanh(log(exp(sum6[0]) + afp(1.f)));
-        sum6[1] = sum6[1] * tanh(log(exp(sum6[1]) + afp(1.f)));
-        sum7[0] = sum7[0] * tanh(log(exp(sum7[0]) + afp(1.f)));
-        sum7[1] = sum7[1] * tanh(log(exp(sum7[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4[0] = sum4[0] * clamp(sum4[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4[1] = sum4[1] * clamp(sum4[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5[0] = sum5[0] * clamp(sum5[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5[1] = sum5[1] * clamp(sum5[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6[0] = sum6[0] * clamp(sum6[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6[1] = sum6[1] * clamp(sum6[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7[0] = sum7[0] * clamp(sum7[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7[1] = sum7[1] * clamp(sum7[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afpvec8(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afpvec8(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afpvec8(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afpvec8(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_pack4to8_1x1s1d1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -222,77 +225,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack4to8_gemm.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -245,77 +248,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -342,125 +345,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-        sum4[0] = max(sum4[0], afp(0.f));
-        sum4[1] = max(sum4[1], afp(0.f));
-        sum5[0] = max(sum5[0], afp(0.f));
-        sum5[1] = max(sum5[1], afp(0.f));
-        sum6[0] = max(sum6[0], afp(0.f));
-        sum6[1] = max(sum6[1], afp(0.f));
-        sum7[0] = max(sum7[0], afp(0.f));
-        sum7[1] = max(sum7[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-        sum4[0] = mix(sum4[0], sum4[0] * afp(slope), lessThan(sum4[0], afpvec4(0.f)));
-        sum4[1] = mix(sum4[1], sum4[1] * afp(slope), lessThan(sum4[1], afpvec4(0.f)));
-        sum5[0] = mix(sum5[0], sum5[0] * afp(slope), lessThan(sum5[0], afpvec4(0.f)));
-        sum5[1] = mix(sum5[1], sum5[1] * afp(slope), lessThan(sum5[1], afpvec4(0.f)));
-        sum6[0] = mix(sum6[0], sum6[0] * afp(slope), lessThan(sum6[0], afpvec4(0.f)));
-        sum6[1] = mix(sum6[1], sum6[1] * afp(slope), lessThan(sum6[1], afpvec4(0.f)));
-        sum7[0] = mix(sum7[0], sum7[0] * afp(slope), lessThan(sum7[0], afpvec4(0.f)));
-        sum7[1] = mix(sum7[1], sum7[1] * afp(slope), lessThan(sum7[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-        sum4[0] = clamp(sum4[0], const_min, const_max);
-        sum4[1] = clamp(sum4[1], const_min, const_max);
-        sum5[0] = clamp(sum5[0], const_min, const_max);
-        sum5[1] = clamp(sum5[1], const_min, const_max);
-        sum6[0] = clamp(sum6[0], const_min, const_max);
-        sum6[1] = clamp(sum6[1], const_min, const_max);
-        sum7[0] = clamp(sum7[0], const_min, const_max);
-        sum7[1] = clamp(sum7[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-        sum4[0] = afp(1.f) / (afp(1.f) + exp(-sum4[0]));
-        sum4[1] = afp(1.f) / (afp(1.f) + exp(-sum4[1]));
-        sum5[0] = afp(1.f) / (afp(1.f) + exp(-sum5[0]));
-        sum5[1] = afp(1.f) / (afp(1.f) + exp(-sum5[1]));
-        sum6[0] = afp(1.f) / (afp(1.f) + exp(-sum6[0]));
-        sum6[1] = afp(1.f) / (afp(1.f) + exp(-sum6[1]));
-        sum7[0] = afp(1.f) / (afp(1.f) + exp(-sum7[0]));
-        sum7[1] = afp(1.f) / (afp(1.f) + exp(-sum7[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-        sum4[0] = sum4[0] * tanh(log(exp(sum4[0]) + afp(1.f)));
-        sum4[1] = sum4[1] * tanh(log(exp(sum4[1]) + afp(1.f)));
-        sum5[0] = sum5[0] * tanh(log(exp(sum5[0]) + afp(1.f)));
-        sum5[1] = sum5[1] * tanh(log(exp(sum5[1]) + afp(1.f)));
-        sum6[0] = sum6[0] * tanh(log(exp(sum6[0]) + afp(1.f)));
-        sum6[1] = sum6[1] * tanh(log(exp(sum6[1]) + afp(1.f)));
-        sum7[0] = sum7[0] * tanh(log(exp(sum7[0]) + afp(1.f)));
-        sum7[1] = sum7[1] * tanh(log(exp(sum7[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4[0] = sum4[0] * clamp(sum4[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4[1] = sum4[1] * clamp(sum4[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5[0] = sum5[0] * clamp(sum5[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5[1] = sum5[1] * clamp(sum5[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6[0] = sum6[0] * clamp(sum6[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6[1] = sum6[1] * clamp(sum6[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7[0] = sum7[0] * clamp(sum7[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7[1] = sum7[1] * clamp(sum7[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afpvec8(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afpvec8(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afpvec8(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afpvec8(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_pack8_1x1s1d1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -222,77 +225,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack8_3x3s1d1_winograd23_transform_output.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -150,77 +153,10 @@ void main()
         v11 = m11 - m12 + m13;
     }
 
-    if (activation_type == 1)
-    {
-        v00[0] = max(v00[0], afp(0.f));
-        v00[1] = max(v00[1], afp(0.f));
-        v10[0] = max(v10[0], afp(0.f));
-        v10[1] = max(v10[1], afp(0.f));
-        v01[0] = max(v01[0], afp(0.f));
-        v01[1] = max(v01[1], afp(0.f));
-        v11[0] = max(v11[0], afp(0.f));
-        v11[1] = max(v11[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        v00[0] = mix(v00[0], v00[0] * afp(slope), lessThan(v00[0], afpvec4(0.f)));
-        v00[1] = mix(v00[1], v00[1] * afp(slope), lessThan(v00[1], afpvec4(0.f)));
-        v10[0] = mix(v10[0], v10[0] * afp(slope), lessThan(v10[0], afpvec4(0.f)));
-        v10[1] = mix(v10[1], v10[1] * afp(slope), lessThan(v10[1], afpvec4(0.f)));
-        v01[0] = mix(v01[0], v01[0] * afp(slope), lessThan(v01[0], afpvec4(0.f)));
-        v01[1] = mix(v01[1], v01[1] * afp(slope), lessThan(v01[1], afpvec4(0.f)));
-        v11[0] = mix(v11[0], v11[0] * afp(slope), lessThan(v11[0], afpvec4(0.f)));
-        v11[1] = mix(v11[1], v11[1] * afp(slope), lessThan(v11[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        v00[0] = clamp(v00[0], const_min, const_max);
-        v00[1] = clamp(v00[1], const_min, const_max);
-        v10[0] = clamp(v10[0], const_min, const_max);
-        v10[1] = clamp(v10[1], const_min, const_max);
-        v01[0] = clamp(v01[0], const_min, const_max);
-        v01[1] = clamp(v01[1], const_min, const_max);
-        v11[0] = clamp(v11[0], const_min, const_max);
-        v11[1] = clamp(v11[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        v00[0] = afp(1.f) / (afp(1.f) + exp(-v00[0]));
-        v00[1] = afp(1.f) / (afp(1.f) + exp(-v00[1]));
-        v10[0] = afp(1.f) / (afp(1.f) + exp(-v10[0]));
-        v10[1] = afp(1.f) / (afp(1.f) + exp(-v10[1]));
-        v01[0] = afp(1.f) / (afp(1.f) + exp(-v01[0]));
-        v01[1] = afp(1.f) / (afp(1.f) + exp(-v01[1]));
-        v11[0] = afp(1.f) / (afp(1.f) + exp(-v11[0]));
-        v11[1] = afp(1.f) / (afp(1.f) + exp(-v11[1]));
-    }
-    if (activation_type == 5)
-    {
-        v00[0] = v00[0] * tanh(log(exp(v00[0]) + afp(1.f)));
-        v00[1] = v00[1] * tanh(log(exp(v00[1]) + afp(1.f)));
-        v10[0] = v10[0] * tanh(log(exp(v10[0]) + afp(1.f)));
-        v10[1] = v10[1] * tanh(log(exp(v10[1]) + afp(1.f)));
-        v01[0] = v01[0] * tanh(log(exp(v01[0]) + afp(1.f)));
-        v01[1] = v01[1] * tanh(log(exp(v01[1]) + afp(1.f)));
-        v11[0] = v11[0] * tanh(log(exp(v11[0]) + afp(1.f)));
-        v11[1] = v11[1] * tanh(log(exp(v11[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        v00[0] = v00[0] * clamp(v00[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v00[1] = v00[1] * clamp(v00[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v01[0] = v01[0] * clamp(v01[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v01[1] = v01[1] * clamp(v01[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v10[0] = v10[0] * clamp(v10[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v10[1] = v10[1] * clamp(v10[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v11[0] = v11[0] * clamp(v11[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v11[1] = v11[1] * clamp(v11[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    v00 = activation_afpvec8(v00, activation_type, activation_param_0, activation_param_1);
+    v01 = activation_afpvec8(v01, activation_type, activation_param_0, activation_param_1);
+    v10 = activation_afpvec8(v10, activation_type, activation_param_0, activation_param_1);
+    v11 = activation_afpvec8(v11, activation_type, activation_param_0, activation_param_1);
 
     // store 2x2
     int x = gx * 2;

src/layer/vulkan/shader/convolution_pack8_3x3s1d1_winograd43_transform_output.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -238,221 +241,22 @@ void main()
         v33 = bias_value + v33;
     }
 
-    if (activation_type == 1)
-    {
-        v00[0] = max(v00[0], afp(0.f));
-        v00[1] = max(v00[1], afp(0.f));
-        v01[0] = max(v01[0], afp(0.f));
-        v01[1] = max(v01[1], afp(0.f));
-        v02[0] = max(v02[0], afp(0.f));
-        v02[1] = max(v02[1], afp(0.f));
-        v03[0] = max(v03[0], afp(0.f));
-        v03[1] = max(v03[1], afp(0.f));
-        v10[0] = max(v10[0], afp(0.f));
-        v10[1] = max(v10[1], afp(0.f));
-        v11[0] = max(v11[0], afp(0.f));
-        v11[1] = max(v11[1], afp(0.f));
-        v12[0] = max(v12[0], afp(0.f));
-        v12[1] = max(v12[1], afp(0.f));
-        v13[0] = max(v13[0], afp(0.f));
-        v13[1] = max(v13[1], afp(0.f));
-        v20[0] = max(v20[0], afp(0.f));
-        v20[1] = max(v20[1], afp(0.f));
-        v21[0] = max(v21[0], afp(0.f));
-        v21[1] = max(v21[1], afp(0.f));
-        v22[0] = max(v22[0], afp(0.f));
-        v22[1] = max(v22[1], afp(0.f));
-        v23[0] = max(v23[0], afp(0.f));
-        v23[1] = max(v23[1], afp(0.f));
-        v30[0] = max(v30[0], afp(0.f));
-        v30[1] = max(v30[1], afp(0.f));
-        v31[0] = max(v31[0], afp(0.f));
-        v31[1] = max(v31[1], afp(0.f));
-        v32[0] = max(v32[0], afp(0.f));
-        v32[1] = max(v32[1], afp(0.f));
-        v33[0] = max(v33[0], afp(0.f));
-        v33[1] = max(v33[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        v00[0] = mix(v00[0], v00[0] * afp(slope), lessThan(v00[0], afpvec4(0.f)));
-        v00[1] = mix(v00[1], v00[1] * afp(slope), lessThan(v00[1], afpvec4(0.f)));
-        v01[0] = mix(v01[0], v01[0] * afp(slope), lessThan(v01[0], afpvec4(0.f)));
-        v01[1] = mix(v01[1], v01[1] * afp(slope), lessThan(v01[1], afpvec4(0.f)));
-        v02[0] = mix(v02[0], v02[0] * afp(slope), lessThan(v02[0], afpvec4(0.f)));
-        v02[1] = mix(v02[1], v02[1] * afp(slope), lessThan(v02[1], afpvec4(0.f)));
-        v03[0] = mix(v03[0], v03[0] * afp(slope), lessThan(v03[0], afpvec4(0.f)));
-        v03[1] = mix(v03[1], v03[1] * afp(slope), lessThan(v03[1], afpvec4(0.f)));
-        v10[0] = mix(v10[0], v10[0] * afp(slope), lessThan(v10[0], afpvec4(0.f)));
-        v10[1] = mix(v10[1], v10[1] * afp(slope), lessThan(v10[1], afpvec4(0.f)));
-        v11[0] = mix(v11[0], v11[0] * afp(slope), lessThan(v11[0], afpvec4(0.f)));
-        v11[1] = mix(v11[1], v11[1] * afp(slope), lessThan(v11[1], afpvec4(0.f)));
-        v12[0] = mix(v12[0], v12[0] * afp(slope), lessThan(v12[0], afpvec4(0.f)));
-        v12[1] = mix(v12[1], v12[1] * afp(slope), lessThan(v12[1], afpvec4(0.f)));
-        v13[0] = mix(v13[0], v13[0] * afp(slope), lessThan(v13[0], afpvec4(0.f)));
-        v13[1] = mix(v13[1], v13[1] * afp(slope), lessThan(v13[1], afpvec4(0.f)));
-        v20[0] = mix(v20[0], v20[0] * afp(slope), lessThan(v20[0], afpvec4(0.f)));
-        v20[1] = mix(v20[1], v20[1] * afp(slope), lessThan(v20[1], afpvec4(0.f)));
-        v21[0] = mix(v21[0], v21[0] * afp(slope), lessThan(v21[0], afpvec4(0.f)));
-        v21[1] = mix(v21[1], v21[1] * afp(slope), lessThan(v21[1], afpvec4(0.f)));
-        v22[0] = mix(v22[0], v22[0] * afp(slope), lessThan(v22[0], afpvec4(0.f)));
-        v22[1] = mix(v22[1], v22[1] * afp(slope), lessThan(v22[1], afpvec4(0.f)));
-        v23[0] = mix(v23[0], v23[0] * afp(slope), lessThan(v23[0], afpvec4(0.f)));
-        v23[1] = mix(v23[1], v23[1] * afp(slope), lessThan(v23[1], afpvec4(0.f)));
-        v30[0] = mix(v30[0], v30[0] * afp(slope), lessThan(v30[0], afpvec4(0.f)));
-        v30[1] = mix(v30[1], v30[1] * afp(slope), lessThan(v30[1], afpvec4(0.f)));
-        v31[0] = mix(v31[0], v31[0] * afp(slope), lessThan(v31[0], afpvec4(0.f)));
-        v31[1] = mix(v31[1], v31[1] * afp(slope), lessThan(v31[1], afpvec4(0.f)));
-        v32[0] = mix(v32[0], v32[0] * afp(slope), lessThan(v32[0], afpvec4(0.f)));
-        v32[1] = mix(v32[1], v32[1] * afp(slope), lessThan(v32[1], afpvec4(0.f)));
-        v33[0] = mix(v33[0], v33[0] * afp(slope), lessThan(v33[0], afpvec4(0.f)));
-        v33[1] = mix(v33[1], v33[1] * afp(slope), lessThan(v33[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        v00[0] = clamp(v00[0], const_min, const_max);
-        v00[1] = clamp(v00[1], const_min, const_max);
-        v01[0] = clamp(v01[0], const_min, const_max);
-        v01[1] = clamp(v01[1], const_min, const_max);
-        v02[0] = clamp(v02[0], const_min, const_max);
-        v02[1] = clamp(v02[1], const_min, const_max);
-        v03[0] = clamp(v03[0], const_min, const_max);
-        v03[1] = clamp(v03[1], const_min, const_max);
-        v10[0] = clamp(v10[0], const_min, const_max);
-        v10[1] = clamp(v10[1], const_min, const_max);
-        v11[0] = clamp(v11[0], const_min, const_max);
-        v11[1] = clamp(v11[1], const_min, const_max);
-        v12[0] = clamp(v12[0], const_min, const_max);
-        v12[1] = clamp(v12[1], const_min, const_max);
-        v13[0] = clamp(v13[0], const_min, const_max);
-        v13[1] = clamp(v13[1], const_min, const_max);
-        v20[0] = clamp(v20[0], const_min, const_max);
-        v20[1] = clamp(v20[1], const_min, const_max);
-        v21[0] = clamp(v21[0], const_min, const_max);
-        v21[1] = clamp(v21[1], const_min, const_max);
-        v22[0] = clamp(v22[0], const_min, const_max);
-        v22[1] = clamp(v22[1], const_min, const_max);
-        v23[0] = clamp(v23[0], const_min, const_max);
-        v23[1] = clamp(v23[1], const_min, const_max);
-        v30[0] = clamp(v30[0], const_min, const_max);
-        v30[1] = clamp(v30[1], const_min, const_max);
-        v31[0] = clamp(v31[0], const_min, const_max);
-        v31[1] = clamp(v31[1], const_min, const_max);
-        v32[0] = clamp(v32[0], const_min, const_max);
-        v32[1] = clamp(v32[1], const_min, const_max);
-        v33[0] = clamp(v33[0], const_min, const_max);
-        v33[1] = clamp(v33[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        v00[0] = afp(1.f) / (afp(1.f) + exp(-v00[0]));
-        v00[1] = afp(1.f) / (afp(1.f) + exp(-v00[1]));
-        v01[0] = afp(1.f) / (afp(1.f) + exp(-v01[0]));
-        v01[1] = afp(1.f) / (afp(1.f) + exp(-v01[1]));
-        v02[0] = afp(1.f) / (afp(1.f) + exp(-v02[0]));
-        v02[1] = afp(1.f) / (afp(1.f) + exp(-v02[1]));
-        v03[0] = afp(1.f) / (afp(1.f) + exp(-v03[0]));
-        v03[1] = afp(1.f) / (afp(1.f) + exp(-v03[1]));
-        v10[0] = afp(1.f) / (afp(1.f) + exp(-v10[0]));
-        v10[1] = afp(1.f) / (afp(1.f) + exp(-v10[1]));
-        v11[0] = afp(1.f) / (afp(1.f) + exp(-v11[0]));
-        v11[1] = afp(1.f) / (afp(1.f) + exp(-v11[1]));
-        v12[0] = afp(1.f) / (afp(1.f) + exp(-v12[0]));
-        v12[1] = afp(1.f) / (afp(1.f) + exp(-v12[1]));
-        v13[0] = afp(1.f) / (afp(1.f) + exp(-v13[0]));
-        v13[1] = afp(1.f) / (afp(1.f) + exp(-v13[1]));
-        v20[0] = afp(1.f) / (afp(1.f) + exp(-v20[0]));
-        v20[1] = afp(1.f) / (afp(1.f) + exp(-v20[1]));
-        v21[0] = afp(1.f) / (afp(1.f) + exp(-v21[0]));
-        v21[1] = afp(1.f) / (afp(1.f) + exp(-v21[1]));
-        v22[0] = afp(1.f) / (afp(1.f) + exp(-v22[0]));
-        v22[1] = afp(1.f) / (afp(1.f) + exp(-v22[1]));
-        v23[0] = afp(1.f) / (afp(1.f) + exp(-v23[0]));
-        v23[1] = afp(1.f) / (afp(1.f) + exp(-v23[1]));
-        v30[0] = afp(1.f) / (afp(1.f) + exp(-v30[0]));
-        v30[1] = afp(1.f) / (afp(1.f) + exp(-v30[1]));
-        v31[0] = afp(1.f) / (afp(1.f) + exp(-v31[0]));
-        v31[1] = afp(1.f) / (afp(1.f) + exp(-v31[1]));
-        v32[0] = afp(1.f) / (afp(1.f) + exp(-v32[0]));
-        v32[1] = afp(1.f) / (afp(1.f) + exp(-v32[1]));
-        v33[0] = afp(1.f) / (afp(1.f) + exp(-v33[0]));
-        v33[1] = afp(1.f) / (afp(1.f) + exp(-v33[1]));
-    }
-    if (activation_type == 5)
-    {
-        v00[0] = v00[0] * tanh(log(exp(v00[0]) + afp(1.f)));
-        v00[1] = v00[1] * tanh(log(exp(v00[1]) + afp(1.f)));
-        v01[0] = v01[0] * tanh(log(exp(v01[0]) + afp(1.f)));
-        v01[1] = v01[1] * tanh(log(exp(v01[1]) + afp(1.f)));
-        v02[0] = v02[0] * tanh(log(exp(v02[0]) + afp(1.f)));
-        v02[1] = v02[1] * tanh(log(exp(v02[1]) + afp(1.f)));
-        v03[0] = v03[0] * tanh(log(exp(v03[0]) + afp(1.f)));
-        v03[1] = v03[1] * tanh(log(exp(v03[1]) + afp(1.f)));
-        v10[0] = v10[0] * tanh(log(exp(v10[0]) + afp(1.f)));
-        v10[1] = v10[1] * tanh(log(exp(v10[1]) + afp(1.f)));
-        v11[0] = v11[0] * tanh(log(exp(v11[0]) + afp(1.f)));
-        v11[1] = v11[1] * tanh(log(exp(v11[1]) + afp(1.f)));
-        v12[0] = v12[0] * tanh(log(exp(v12[0]) + afp(1.f)));
-        v12[1] = v12[1] * tanh(log(exp(v12[1]) + afp(1.f)));
-        v13[0] = v13[0] * tanh(log(exp(v13[0]) + afp(1.f)));
-        v13[1] = v13[1] * tanh(log(exp(v13[1]) + afp(1.f)));
-        v20[0] = v20[0] * tanh(log(exp(v20[0]) + afp(1.f)));
-        v20[1] = v20[1] * tanh(log(exp(v20[1]) + afp(1.f)));
-        v21[0] = v21[0] * tanh(log(exp(v21[0]) + afp(1.f)));
-        v21[1] = v21[1] * tanh(log(exp(v21[1]) + afp(1.f)));
-        v22[0] = v22[0] * tanh(log(exp(v22[0]) + afp(1.f)));
-        v22[1] = v22[1] * tanh(log(exp(v22[1]) + afp(1.f)));
-        v23[0] = v23[0] * tanh(log(exp(v23[0]) + afp(1.f)));
-        v23[1] = v23[1] * tanh(log(exp(v23[1]) + afp(1.f)));
-        v30[0] = v30[0] * tanh(log(exp(v30[0]) + afp(1.f)));
-        v30[1] = v30[1] * tanh(log(exp(v30[1]) + afp(1.f)));
-        v31[0] = v31[0] * tanh(log(exp(v31[0]) + afp(1.f)));
-        v31[1] = v31[1] * tanh(log(exp(v31[1]) + afp(1.f)));
-        v32[0] = v32[0] * tanh(log(exp(v32[0]) + afp(1.f)));
-        v32[1] = v32[1] * tanh(log(exp(v32[1]) + afp(1.f)));
-        v33[0] = v33[0] * tanh(log(exp(v33[0]) + afp(1.f)));
-        v33[1] = v33[1] * tanh(log(exp(v33[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        v00[0] = v00[0] * clamp(v00[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v00[1] = v00[1] * clamp(v00[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v01[0] = v01[0] * clamp(v01[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v01[1] = v01[1] * clamp(v01[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v02[0] = v02[0] * clamp(v02[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v02[1] = v02[1] * clamp(v02[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v03[0] = v03[0] * clamp(v03[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v03[1] = v03[1] * clamp(v03[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v10[0] = v10[0] * clamp(v10[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v10[1] = v10[1] * clamp(v10[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v11[0] = v11[0] * clamp(v11[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v11[1] = v11[1] * clamp(v11[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v12[0] = v12[0] * clamp(v12[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v12[1] = v12[1] * clamp(v12[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v13[0] = v13[0] * clamp(v13[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v13[1] = v13[1] * clamp(v13[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v20[0] = v20[0] * clamp(v20[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v20[1] = v20[1] * clamp(v20[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v21[0] = v21[0] * clamp(v21[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v21[1] = v21[1] * clamp(v21[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v22[0] = v22[0] * clamp(v22[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v22[1] = v22[1] * clamp(v22[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v23[0] = v23[0] * clamp(v23[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v23[1] = v23[1] * clamp(v23[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v30[0] = v30[0] * clamp(v30[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v30[1] = v30[1] * clamp(v30[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v31[0] = v31[0] * clamp(v31[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v31[1] = v31[1] * clamp(v31[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v32[0] = v32[0] * clamp(v32[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v32[1] = v32[1] * clamp(v32[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v33[0] = v33[0] * clamp(v33[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        v33[1] = v33[1] * clamp(v33[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    v00 = activation_afpvec8(v00, activation_type, activation_param_0, activation_param_1);
+    v01 = activation_afpvec8(v01, activation_type, activation_param_0, activation_param_1);
+    v02 = activation_afpvec8(v02, activation_type, activation_param_0, activation_param_1);
+    v03 = activation_afpvec8(v03, activation_type, activation_param_0, activation_param_1);
+    v10 = activation_afpvec8(v10, activation_type, activation_param_0, activation_param_1);
+    v11 = activation_afpvec8(v11, activation_type, activation_param_0, activation_param_1);
+    v12 = activation_afpvec8(v12, activation_type, activation_param_0, activation_param_1);
+    v13 = activation_afpvec8(v13, activation_type, activation_param_0, activation_param_1);
+    v20 = activation_afpvec8(v20, activation_type, activation_param_0, activation_param_1);
+    v21 = activation_afpvec8(v21, activation_type, activation_param_0, activation_param_1);
+    v22 = activation_afpvec8(v22, activation_type, activation_param_0, activation_param_1);
+    v23 = activation_afpvec8(v23, activation_type, activation_param_0, activation_param_1);
+    v30 = activation_afpvec8(v30, activation_type, activation_param_0, activation_param_1);
+    v31 = activation_afpvec8(v31, activation_type, activation_param_0, activation_param_1);
+    v32 = activation_afpvec8(v32, activation_type, activation_param_0, activation_param_1);
+    v33 = activation_afpvec8(v33, activation_type, activation_param_0, activation_param_1);
 
     // store 4x4
     int x = gx * 4;

src/layer/vulkan/shader/convolution_pack8_gemm.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -245,77 +248,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0[0] = max(sum0[0], afp(0.f));
-        sum0[1] = max(sum0[1], afp(0.f));
-        sum1[0] = max(sum1[0], afp(0.f));
-        sum1[1] = max(sum1[1], afp(0.f));
-        sum2[0] = max(sum2[0], afp(0.f));
-        sum2[1] = max(sum2[1], afp(0.f));
-        sum3[0] = max(sum3[0], afp(0.f));
-        sum3[1] = max(sum3[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0[0] = mix(sum0[0], sum0[0] * afp(slope), lessThan(sum0[0], afpvec4(0.f)));
-        sum0[1] = mix(sum0[1], sum0[1] * afp(slope), lessThan(sum0[1], afpvec4(0.f)));
-        sum1[0] = mix(sum1[0], sum1[0] * afp(slope), lessThan(sum1[0], afpvec4(0.f)));
-        sum1[1] = mix(sum1[1], sum1[1] * afp(slope), lessThan(sum1[1], afpvec4(0.f)));
-        sum2[0] = mix(sum2[0], sum2[0] * afp(slope), lessThan(sum2[0], afpvec4(0.f)));
-        sum2[1] = mix(sum2[1], sum2[1] * afp(slope), lessThan(sum2[1], afpvec4(0.f)));
-        sum3[0] = mix(sum3[0], sum3[0] * afp(slope), lessThan(sum3[0], afpvec4(0.f)));
-        sum3[1] = mix(sum3[1], sum3[1] * afp(slope), lessThan(sum3[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0[0] = clamp(sum0[0], const_min, const_max);
-        sum0[1] = clamp(sum0[1], const_min, const_max);
-        sum1[0] = clamp(sum1[0], const_min, const_max);
-        sum1[1] = clamp(sum1[1], const_min, const_max);
-        sum2[0] = clamp(sum2[0], const_min, const_max);
-        sum2[1] = clamp(sum2[1], const_min, const_max);
-        sum3[0] = clamp(sum3[0], const_min, const_max);
-        sum3[1] = clamp(sum3[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0[0] = afp(1.f) / (afp(1.f) + exp(-sum0[0]));
-        sum0[1] = afp(1.f) / (afp(1.f) + exp(-sum0[1]));
-        sum1[0] = afp(1.f) / (afp(1.f) + exp(-sum1[0]));
-        sum1[1] = afp(1.f) / (afp(1.f) + exp(-sum1[1]));
-        sum2[0] = afp(1.f) / (afp(1.f) + exp(-sum2[0]));
-        sum2[1] = afp(1.f) / (afp(1.f) + exp(-sum2[1]));
-        sum3[0] = afp(1.f) / (afp(1.f) + exp(-sum3[0]));
-        sum3[1] = afp(1.f) / (afp(1.f) + exp(-sum3[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum0[0] = sum0[0] * tanh(log(exp(sum0[0]) + afp(1.f)));
-        sum0[1] = sum0[1] * tanh(log(exp(sum0[1]) + afp(1.f)));
-        sum1[0] = sum1[0] * tanh(log(exp(sum1[0]) + afp(1.f)));
-        sum1[1] = sum1[1] * tanh(log(exp(sum1[1]) + afp(1.f)));
-        sum2[0] = sum2[0] * tanh(log(exp(sum2[0]) + afp(1.f)));
-        sum2[1] = sum2[1] * tanh(log(exp(sum2[1]) + afp(1.f)));
-        sum3[0] = sum3[0] * tanh(log(exp(sum3[0]) + afp(1.f)));
-        sum3[1] = sum3[1] * tanh(log(exp(sum3[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0[0] = sum0[0] * clamp(sum0[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum0[1] = sum0[1] * clamp(sum0[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[0] = sum1[0] * clamp(sum1[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1[1] = sum1[1] * clamp(sum1[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[0] = sum2[0] * clamp(sum2[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2[1] = sum2[1] * clamp(sum2[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[0] = sum3[0] * clamp(sum3[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3[1] = sum3[1] * clamp(sum3[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack8to1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -198,77 +201,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-        sum4 = max(sum4, afp(0.f));
-        sum5 = max(sum5, afp(0.f));
-        sum6 = max(sum6, afp(0.f));
-        sum7 = max(sum7, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = sum0 < afp(0.f) ? sum0 * slope : sum0;
-        sum1 = sum1 < afp(0.f) ? sum1 * slope : sum1;
-        sum2 = sum2 < afp(0.f) ? sum2 * slope : sum2;
-        sum3 = sum3 < afp(0.f) ? sum3 * slope : sum3;
-        sum4 = sum4 < afp(0.f) ? sum4 * slope : sum4;
-        sum5 = sum5 < afp(0.f) ? sum5 * slope : sum5;
-        sum6 = sum6 < afp(0.f) ? sum6 * slope : sum6;
-        sum7 = sum7 < afp(0.f) ? sum7 * slope : sum7;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-        sum4 = clamp(sum4, const_min, const_max);
-        sum5 = clamp(sum5, const_min, const_max);
-        sum6 = clamp(sum6, const_min, const_max);
-        sum7 = clamp(sum7, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-        sum4 = afp(1.f) / (afp(1.f) + exp(-sum4));
-        sum5 = afp(1.f) / (afp(1.f) + exp(-sum5));
-        sum6 = afp(1.f) / (afp(1.f) + exp(-sum6));
-        sum7 = afp(1.f) / (afp(1.f) + exp(-sum7));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-        sum4 = sum4 * tanh(log(exp(sum4) + afp(1.f)));
-        sum5 = sum5 * tanh(log(exp(sum5) + afp(1.f)));
-        sum6 = sum6 * tanh(log(exp(sum6) + afp(1.f)));
-        sum7 = sum7 * tanh(log(exp(sum7) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4 = sum4 * clamp(sum4 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5 = sum5 * clamp(sum5 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6 = sum6 * clamp(sum6 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7 = sum7 * clamp(sum7 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afp(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afp(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afp(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afp(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_pack8to1_1x1s1d1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -146,53 +149,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = sum0 < afp(0.f) ? sum0 * slope : sum0;
-        sum1 = sum1 < afp(0.f) ? sum1 * slope : sum1;
-        sum2 = sum2 < afp(0.f) ? sum2 * slope : sum2;
-        sum3 = sum3 < afp(0.f) ? sum3 * slope : sum3;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack8to1_gemm.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -169,53 +172,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = sum0 < afp(0.f) ? sum0 * slope : sum0;
-        sum1 = sum1 < afp(0.f) ? sum1 * slope : sum1;
-        sum2 = sum2 < afp(0.f) ? sum2 * slope : sum2;
-        sum3 = sum3 < afp(0.f) ? sum3 * slope : sum3;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack8to4.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -258,77 +261,14 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-        sum4 = max(sum4, afp(0.f));
-        sum5 = max(sum5, afp(0.f));
-        sum6 = max(sum6, afp(0.f));
-        sum7 = max(sum7, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-        sum4 = mix(sum4, sum4 * afp(slope), lessThan(sum4, afpvec4(0.f)));
-        sum5 = mix(sum5, sum5 * afp(slope), lessThan(sum5, afpvec4(0.f)));
-        sum6 = mix(sum6, sum6 * afp(slope), lessThan(sum6, afpvec4(0.f)));
-        sum7 = mix(sum7, sum7 * afp(slope), lessThan(sum7, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-        sum4 = clamp(sum4, const_min, const_max);
-        sum5 = clamp(sum5, const_min, const_max);
-        sum6 = clamp(sum6, const_min, const_max);
-        sum7 = clamp(sum7, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-        sum4 = afp(1.f) / (afp(1.f) + exp(-sum4));
-        sum5 = afp(1.f) / (afp(1.f) + exp(-sum5));
-        sum6 = afp(1.f) / (afp(1.f) + exp(-sum6));
-        sum7 = afp(1.f) / (afp(1.f) + exp(-sum7));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-        sum4 = sum4 * tanh(log(exp(sum4) + afp(1.f)));
-        sum5 = sum5 * tanh(log(exp(sum5) + afp(1.f)));
-        sum6 = sum6 * tanh(log(exp(sum6) + afp(1.f)));
-        sum7 = sum7 * tanh(log(exp(sum7) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum4 = sum4 * clamp(sum4 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum5 = sum5 * clamp(sum5 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum6 = sum6 * clamp(sum6 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum7 = sum7 * clamp(sum7 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
+    sum4 = activation_afpvec4(sum4, activation_type, activation_param_0, activation_param_1);
+    sum5 = activation_afpvec4(sum5, activation_type, activation_param_0, activation_param_1);
+    sum6 = activation_afpvec4(sum6, activation_type, activation_param_0, activation_param_1);
+    sum7 = activation_afpvec4(sum7, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);

src/layer/vulkan/shader/convolution_pack8to4_1x1s1d1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -182,53 +185,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolution_pack8to4_gemm.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -205,53 +208,10 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum0 = max(sum0, afp(0.f));
-        sum1 = max(sum1, afp(0.f));
-        sum2 = max(sum2, afp(0.f));
-        sum3 = max(sum3, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum0 = mix(sum0, sum0 * afp(slope), lessThan(sum0, afpvec4(0.f)));
-        sum1 = mix(sum1, sum1 * afp(slope), lessThan(sum1, afpvec4(0.f)));
-        sum2 = mix(sum2, sum2 * afp(slope), lessThan(sum2, afpvec4(0.f)));
-        sum3 = mix(sum3, sum3 * afp(slope), lessThan(sum3, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum0 = clamp(sum0, const_min, const_max);
-        sum1 = clamp(sum1, const_min, const_max);
-        sum2 = clamp(sum2, const_min, const_max);
-        sum3 = clamp(sum3, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum0 = afp(1.f) / (afp(1.f) + exp(-sum0));
-        sum1 = afp(1.f) / (afp(1.f) + exp(-sum1));
-        sum2 = afp(1.f) / (afp(1.f) + exp(-sum2));
-        sum3 = afp(1.f) / (afp(1.f) + exp(-sum3));
-    }
-    if (activation_type == 5)
-    {
-        sum0 = sum0 * tanh(log(exp(sum0) + afp(1.f)));
-        sum1 = sum1 * tanh(log(exp(sum1) + afp(1.f)));
-        sum2 = sum2 * tanh(log(exp(sum2) + afp(1.f)));
-        sum3 = sum3 * tanh(log(exp(sum3) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum0 = sum0 * clamp(sum0 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum1 = sum1 * clamp(sum1 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum2 = sum2 * clamp(sum2 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum3 = sum3 * clamp(sum3 * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
+    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
+    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
+    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);

src/layer/vulkan/shader/convolutiondepthwise.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -132,35 +135,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -152,35 +155,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group_pack1to4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -160,35 +163,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group_pack1to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -165,41 +168,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group_pack4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -180,35 +183,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group_pack4to1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -160,35 +163,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group_pack4to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -191,41 +194,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group_pack8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -191,41 +194,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group_pack8to1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -163,35 +166,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_group_pack8to4.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -175,35 +178,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_pack4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -140,35 +143,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/convolutiondepthwise_pack8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -145,41 +148,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -165,25 +168,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_col2im.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -120,25 +123,7 @@ void main()
         }
     }
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack1to4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -173,25 +176,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack1to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -178,29 +181,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -193,25 +196,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack4_col2im.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -120,25 +123,7 @@ void main()
         }
     }
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack4to1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -173,25 +176,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack4to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -204,29 +207,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -204,29 +207,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack8_col2im.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -121,29 +124,7 @@ void main()
         }
     }
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack8to1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -176,25 +179,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolution_pack8to4.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -188,25 +191,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -159,25 +162,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -177,25 +180,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group_pack1to4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -185,25 +188,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group_pack1to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -190,29 +193,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group_pack4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -205,25 +208,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group_pack4to1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -185,25 +188,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group_pack4to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -216,29 +219,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -216,29 +219,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8to1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -188,25 +191,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8to4.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -200,25 +203,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_pack4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -167,25 +170,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/deconvolutiondepthwise_pack8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int kernel_w = 1;
 layout (constant_id = 1) const int kernel_h = 1;
 layout (constant_id = 2) const int dilation_w = 1;
@@ -172,29 +175,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, gy, gz), sum);

src/layer/vulkan/shader/innerproduct.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -104,35 +107,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_gemm.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -105,35 +108,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, 0), sum);

src/layer/vulkan/shader/innerproduct_gemm_wp1to4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -113,35 +116,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx * 4 + 0, gy, 0), sum.r);

src/layer/vulkan/shader/innerproduct_gemm_wp1to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -118,41 +121,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx * 8 + 0, gy, 0), sum[0].r);

src/layer/vulkan/shader/innerproduct_gemm_wp4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -141,35 +144,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx * 4 + 0, gy, 0), sum.r);

src/layer/vulkan/shader/innerproduct_gemm_wp4to1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -121,35 +124,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, 0), sum);

src/layer/vulkan/shader/innerproduct_gemm_wp4to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -152,41 +155,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx * 8 + 0, gy, 0), sum[0].r);

src/layer/vulkan/shader/innerproduct_gemm_wp8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -160,41 +163,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx * 8 + 0, gy, 0), sum[0].r);

src/layer/vulkan/shader/innerproduct_gemm_wp8to1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -132,35 +135,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, gy, 0), sum);

src/layer/vulkan/shader/innerproduct_gemm_wp8to4.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -144,35 +147,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx * 4 + 0, gy, 0), sum.r);

src/layer/vulkan/shader/innerproduct_pack1to4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -112,35 +115,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_pack1to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -119,41 +122,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_pack4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -135,35 +138,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_pack4to1.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -112,35 +115,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_pack4to8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -147,41 +150,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_pack8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -147,41 +150,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_pack8to1.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -115,35 +118,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_pack8to4.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -131,35 +134,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_reduce_sum8.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -88,35 +91,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = sum < afp(0.f) ? sum * slope : sum;
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_reduce_sum8_pack4.comp

@@ -21,6 +21,9 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -88,35 +91,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum = max(sum, afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum = clamp(sum, const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum = afp(1.f) / (afp(1.f) + exp(-sum));
-    }
-    if (activation_type == 5)
-    {
-        sum = sum * tanh(log(exp(sum) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum = sum * clamp(sum * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/innerproduct_reduce_sum8_pack8.comp

@@ -22,6 +22,9 @@ struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
+#extension GL_GOOGLE_include_directive: enable
+#include "vulkan_activation.comp"
+
 layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
@@ -89,41 +92,7 @@ void main()
     }
 #endif
 
-    if (activation_type == 1)
-    {
-        sum[0] = max(sum[0], afp(0.f));
-        sum[1] = max(sum[1], afp(0.f));
-    }
-    if (activation_type == 2)
-    {
-        const afp slope = afp(activation_param_0);
-        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
-        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
-    }
-    if (activation_type == 3)
-    {
-        const afp const_min = afp(activation_param_0);
-        const afp const_max = afp(activation_param_1);
-        sum[0] = clamp(sum[0], const_min, const_max);
-        sum[1] = clamp(sum[1], const_min, const_max);
-    }
-    if (activation_type == 4)
-    {
-        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
-        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
-    }
-    if (activation_type == 5)
-    {
-        sum[0] = sum[0] * tanh(log(exp(sum[0]) + afp(1.f)));
-        sum[1] = sum[1] * tanh(log(exp(sum[1]) + afp(1.f)));
-    }
-    if (activation_type == 6)
-    {
-        const afp alpha = afp(activation_param_0);
-        const afp beta = afp(activation_param_1);
-        sum[0] = sum[0] * clamp(sum[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-        sum[1] = sum[1] * clamp(sum[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    }
+    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx, 0, 0), sum);

src/layer/vulkan/shader/mish.comp

@@ -61,7 +61,11 @@ void main()
     afp v = buffer_ld1(bottom_top_blob_data, gi);
 #endif
 
+#if NCNN_moltenvk
+    v = v * afp(tanh(float(log(exp(v) + afp(1.f)))));
+#else
     v = v * tanh(log(exp(v) + afp(1.f)));
+#endif
 
 #if NCNN_image_shader
     image3d_st1(top_blob_3d, ivec3(gx, gy, gz), v);

src/layer/vulkan/shader/mish_pack4.comp

@@ -61,7 +61,11 @@ void main()
     afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 #endif
 
+#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
 
 #if NCNN_image_shader
     image3d_st4(top_blob_3d, ivec3(gx, gy, gz), v);

src/layer/vulkan/shader/mish_pack8.comp

@@ -62,8 +62,13 @@ void main()
     afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
 #endif
 
+#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
 
 #if NCNN_image_shader
     image3d_st8(top_blob_3d, ivec3(gx, gy, gz), v);

src/layer/vulkan/shader/tanh.comp

@@ -61,7 +61,11 @@ void main()
     afp v = buffer_ld1(bottom_top_blob_data, gi);
 #endif
 
+#if NCNN_moltenvk
+    v = afp(tanh(float(v)));
+#else
     v = tanh(v);
+#endif
 
 #if NCNN_image_shader
     image3d_st1(top_blob_3d, ivec3(gx, gy, gz), v);

src/layer/vulkan/shader/tanh_pack4.comp

@@ -61,7 +61,11 @@ void main()
     afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 #endif
 
+#if NCNN_moltenvk
+    v = afpvec4(tanh(vec4(v)));
+#else
     v = tanh(v);
+#endif
 
 #if NCNN_image_shader
     image3d_st4(top_blob_3d, ivec3(gx, gy, gz), v);