add option to upload model function, pipeline creation honors option use flags, setting allocator per extractor do not make much sense

src/command.cpp

@@ -17,6 +17,7 @@
 #if NCNN_VULKAN
 
 #include <stdio.h>
+#include "option.h"
 
 namespace ncnn {
 
@@ -841,16 +842,16 @@ VkTransfer::~VkTransfer()
 {
 }
 
-void VkTransfer::record_upload(const Mat& src, VkMat& dst)
+void VkTransfer::record_upload(const Mat& src, VkMat& dst, const Option& opt)
 {
     if (src.elemsize / src.packing == 4)
     {
-        if (vkdev->info.support_fp16_storage || (vkdev->info.support_fp16_packed && src.packing % 4 == 0))
+        if (opt.use_fp16_storage || (opt.use_fp16_packed && src.packing % 4 == 0))
         {
             Mat src_fp16;
             cast_float32_to_float16(src, src_fp16);
 
-            record_upload(src_fp16, dst);
+            record_upload(src_fp16, dst, opt);
 
             return;
         }

src/command.h

@@ -173,7 +173,7 @@ public:
     VkTransfer(const VulkanDevice* vkdev);
     ~VkTransfer();
 
-    void record_upload(const Mat& src, VkMat& dst);
+    void record_upload(const Mat& src, VkMat& dst, const Option& opt);
 
     int submit_and_wait();
 

src/layer.cpp

@@ -101,7 +101,7 @@ int Layer::forward_inplace(Mat& /*bottom_top_blob*/, const Option& /*opt*/) cons
 }
 
 #if NCNN_VULKAN
-int Layer::upload_model(VkTransfer& /*cmd*/)
+int Layer::upload_model(VkTransfer& /*cmd*/, const Option& /*opt*/)
 {
     return 0;
 }

src/layer.h

@@ -79,7 +79,7 @@ public:
 #if NCNN_VULKAN
 public:
     // upload weight blob from host to device
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt = Option());
 
 public:
     // implement inference

src/layer/vulkan/absval_vulkan.cpp

@@ -34,14 +34,14 @@ int AbsVal_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_absval = new Pipeline(vkdev);
         pipeline_absval->set_optimal_local_size_xyz();
-        pipeline_absval->create("absval", specializations, 1, 5);
+        pipeline_absval->create("absval", opt, specializations, 1, 5);
     }
 
     // pack4
     {
         pipeline_absval_pack4 = new Pipeline(vkdev);
         pipeline_absval_pack4->set_optimal_local_size_xyz();
-        pipeline_absval_pack4->create("absval_pack4", specializations, 1, 5);
+        pipeline_absval_pack4->create("absval_pack4", opt, specializations, 1, 5);
     }
 
     return 0;

src/layer/vulkan/batchnorm_vulkan.cpp

@@ -37,7 +37,7 @@ int BatchNorm_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_batchnorm = new Pipeline(vkdev);
         pipeline_batchnorm->set_optimal_local_size_xyz(32, 32, channels);
-        pipeline_batchnorm->create("batchnorm", specializations, 3, 5);
+        pipeline_batchnorm->create("batchnorm", opt, specializations, 3, 5);
     }
 
     // pack4
@@ -45,7 +45,7 @@ int BatchNorm_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_batchnorm_pack4 = new Pipeline(vkdev);
         pipeline_batchnorm_pack4->set_optimal_local_size_xyz(32, 32, channels / 4);
-        pipeline_batchnorm_pack4->create("batchnorm_pack4", specializations, 3, 5);
+        pipeline_batchnorm_pack4->create("batchnorm_pack4", opt, specializations, 3, 5);
     }
 
     return 0;
@@ -62,13 +62,13 @@ int BatchNorm_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int BatchNorm_vulkan::upload_model(VkTransfer& cmd)
+int BatchNorm_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
     // pack1
     if (channels % 4 != 0)
     {
-        cmd.record_upload(a_data, a_data_gpu);
-        cmd.record_upload(b_data, b_data_gpu);
+        cmd.record_upload(a_data, a_data_gpu, opt);
+        cmd.record_upload(b_data, b_data_gpu, opt);
     }
 
     // pack4
@@ -76,11 +76,11 @@ int BatchNorm_vulkan::upload_model(VkTransfer& cmd)
     {
         Mat a_data_pack4;
         convert_packing(a_data, a_data_pack4, 4);
-        cmd.record_upload(a_data_pack4, a_data_gpu_pack4);
+        cmd.record_upload(a_data_pack4, a_data_gpu_pack4, opt);
 
         Mat b_data_pack4;
         convert_packing(b_data, b_data_pack4, 4);
-        cmd.record_upload(b_data_pack4, b_data_gpu_pack4);
+        cmd.record_upload(b_data_pack4, b_data_gpu_pack4, opt);
     }
 
     return 0;

src/layer/vulkan/batchnorm_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
 

src/layer/vulkan/binaryop_vulkan.cpp

@@ -40,14 +40,14 @@ int BinaryOp_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_binaryop = new Pipeline(vkdev);
         pipeline_binaryop->set_optimal_local_size_xyz();
-        pipeline_binaryop->create("binaryop", specializations, 3, 15);
+        pipeline_binaryop->create("binaryop", opt, specializations, 3, 15);
     }
 
     // pack4
     {
         pipeline_binaryop_pack4 = new Pipeline(vkdev);
         pipeline_binaryop_pack4->set_optimal_local_size_xyz();
-        pipeline_binaryop_pack4->create("binaryop_pack4", specializations, 3, 15);
+        pipeline_binaryop_pack4->create("binaryop_pack4", opt, specializations, 3, 15);
     }
 
     return 0;

src/layer/vulkan/cast_vulkan.cpp

@@ -38,14 +38,14 @@ int Cast_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_cast_fp32_to_fp16 = new Pipeline(vkdev);
             pipeline_cast_fp32_to_fp16->set_optimal_local_size_xyz();
-            pipeline_cast_fp32_to_fp16->create("cast_fp32_to_fp16", specializations, 2, 10);
+            pipeline_cast_fp32_to_fp16->create("cast_fp32_to_fp16", opt, specializations, 2, 10);
         }
 
         // pack4
         {
             pipeline_cast_fp32_to_fp16_pack4 = new Pipeline(vkdev);
             pipeline_cast_fp32_to_fp16_pack4->set_optimal_local_size_xyz();
-            pipeline_cast_fp32_to_fp16_pack4->create("cast_fp32_to_fp16_pack4", specializations, 2, 10);
+            pipeline_cast_fp32_to_fp16_pack4->create("cast_fp32_to_fp16_pack4", opt, specializations, 2, 10);
         }
     }
 
@@ -55,14 +55,14 @@ int Cast_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_cast_fp16_to_fp32 = new Pipeline(vkdev);
             pipeline_cast_fp16_to_fp32->set_optimal_local_size_xyz();
-            pipeline_cast_fp16_to_fp32->create("cast_fp16_to_fp32", specializations, 2, 10);
+            pipeline_cast_fp16_to_fp32->create("cast_fp16_to_fp32", opt, specializations, 2, 10);
         }
 
         // pack4
         {
             pipeline_cast_fp16_to_fp32_pack4 = new Pipeline(vkdev);
             pipeline_cast_fp16_to_fp32_pack4->set_optimal_local_size_xyz();
-            pipeline_cast_fp16_to_fp32_pack4->create("cast_fp16_to_fp32_pack4", specializations, 2, 10);
+            pipeline_cast_fp16_to_fp32_pack4->create("cast_fp16_to_fp32_pack4", opt, specializations, 2, 10);
         }
     }
 

src/layer/vulkan/clip_vulkan.cpp

@@ -36,14 +36,14 @@ int Clip_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_clip = new Pipeline(vkdev);
         pipeline_clip->set_optimal_local_size_xyz();
-        pipeline_clip->create("clip", specializations, 1, 5);
+        pipeline_clip->create("clip", opt, specializations, 1, 5);
     }
 
     // pack4
     {
         pipeline_clip_pack4 = new Pipeline(vkdev);
         pipeline_clip_pack4->set_optimal_local_size_xyz();
-        pipeline_clip_pack4->create("clip_pack4", specializations, 1, 5);
+        pipeline_clip_pack4->create("clip_pack4", opt, specializations, 1, 5);
     }
 
     return 0;

src/layer/vulkan/concat_vulkan.cpp

@@ -40,30 +40,30 @@ int Concat_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_concat[0] = new Pipeline(vkdev);
         pipeline_concat[0]->set_optimal_local_size_xyz();
-        pipeline_concat[0]->create("concat", specializations, 2, 11);
+        pipeline_concat[0]->create("concat", opt, specializations, 2, 11);
         pipeline_concat[1] = new Pipeline(vkdev);
         pipeline_concat[1]->set_optimal_local_size_xyz();
-        pipeline_concat[1]->create("concat", specializations, 2, 11);
+        pipeline_concat[1]->create("concat", opt, specializations, 2, 11);
     }
 
     // pack4
     {
         pipeline_concat_pack4[0] = new Pipeline(vkdev);
         pipeline_concat_pack4[0]->set_optimal_local_size_xyz();
-        pipeline_concat_pack4[0]->create("concat_pack4", specializations, 2, 11);
+        pipeline_concat_pack4[0]->create("concat_pack4", opt, specializations, 2, 11);
         pipeline_concat_pack4[1] = new Pipeline(vkdev);
         pipeline_concat_pack4[1]->set_optimal_local_size_xyz();
-        pipeline_concat_pack4[1]->create("concat_pack4", specializations, 2, 11);
+        pipeline_concat_pack4[1]->create("concat_pack4", opt, specializations, 2, 11);
     }
 
     // pack4to1
     {
         pipeline_concat_pack4to1[0] = new Pipeline(vkdev);
         pipeline_concat_pack4to1[0]->set_optimal_local_size_xyz();
-        pipeline_concat_pack4to1[0]->create("concat_pack4to1", specializations, 2, 11);
+        pipeline_concat_pack4to1[0]->create("concat_pack4to1", opt, specializations, 2, 11);
         pipeline_concat_pack4to1[1] = new Pipeline(vkdev);
         pipeline_concat_pack4to1[1]->set_optimal_local_size_xyz();
-        pipeline_concat_pack4to1[1]->create("concat_pack4to1", specializations, 2, 11);
+        pipeline_concat_pack4to1[1]->create("concat_pack4to1", opt, specializations, 2, 11);
     }
 
     return 0;

src/layer/vulkan/convolution_vulkan.cpp

@@ -84,7 +84,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_convolution = new Pipeline(vkdev);
         pipeline_convolution->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_convolution->create("convolution", specializations, 4, 10);
+        pipeline_convolution->create("convolution", opt, specializations, 4, 10);
 
         if (kernel_w == 1 && kernel_h == 1 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1)
         {
@@ -97,7 +97,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
             specializations[2].f = activation_params.w == 1 ? activation_params[0] : 0.f;
             specializations[3].f = activation_params.w == 2 ? activation_params[1] : 0.f;
 
-            pipeline_convolution_1x1s1d1->create("convolution_1x1s1d1", specializations, 4, 8);
+            pipeline_convolution_1x1s1d1->create("convolution_1x1s1d1", opt, specializations, 4, 8);
         }
     }
 
@@ -106,7 +106,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_convolution_pack4 = new Pipeline(vkdev);
         pipeline_convolution_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_convolution_pack4->create("convolution_pack4", specializations, 4, 10);
+        pipeline_convolution_pack4->create("convolution_pack4", opt, specializations, 4, 10);
 
         if (kernel_w == 1 && kernel_h == 1 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1)
         {
@@ -119,7 +119,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
             specializations[2].f = activation_params.w == 1 ? activation_params[0] : 0.f;
             specializations[3].f = activation_params.w == 2 ? activation_params[1] : 0.f;
 
-            pipeline_convolution_pack4_1x1s1d1->create("convolution_pack4_1x1s1d1", specializations, 4, 8);
+            pipeline_convolution_pack4_1x1s1d1->create("convolution_pack4_1x1s1d1", opt, specializations, 4, 8);
         }
 
         if (kernel_w == 3 && kernel_h == 3 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1)
@@ -132,7 +132,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
 
             pipeline_convolution_pack4_3x3s1d1_lds_8_8_2 = new Pipeline(vkdev);
             pipeline_convolution_pack4_3x3s1d1_lds_8_8_2->set_local_size_xyz(8, 8, 2);
-            pipeline_convolution_pack4_3x3s1d1_lds_8_8_2->create("convolution_pack4_3x3s1d1_lds_8_8_2", specializations, 4, 10);
+            pipeline_convolution_pack4_3x3s1d1_lds_8_8_2->create("convolution_pack4_3x3s1d1_lds_8_8_2", opt, specializations, 4, 10);
 
             if (num_input >= 16 && num_output >= 16)
             {
@@ -172,15 +172,15 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
 
                 pipeline_convolution_pack4_3x3s1d1_winograd23_transform_input = new Pipeline(vkdev);
                 pipeline_convolution_pack4_3x3s1d1_winograd23_transform_input->set_local_size_xyz(8, 8, 1);
-                pipeline_convolution_pack4_3x3s1d1_winograd23_transform_input->create("convolution_pack4_3x3s1d1_winograd23_transform_input", std::vector<vk_specialization_type>(), 2, 7);
+                pipeline_convolution_pack4_3x3s1d1_winograd23_transform_input->create("convolution_pack4_3x3s1d1_winograd23_transform_input", opt, std::vector<vk_specialization_type>(), 2, 7);
 
                 pipeline_convolution_pack4_3x3s1d1_winograd23_gemm = new Pipeline(vkdev);
                 pipeline_convolution_pack4_3x3s1d1_winograd23_gemm->set_local_size_xyz(4, 4, 4);
-                pipeline_convolution_pack4_3x3s1d1_winograd23_gemm->create("convolution_pack4_3x3s1d1_winograd23_gemm", std::vector<vk_specialization_type>(), 3, 6);
+                pipeline_convolution_pack4_3x3s1d1_winograd23_gemm->create("convolution_pack4_3x3s1d1_winograd23_gemm", opt, std::vector<vk_specialization_type>(), 3, 6);
 
                 pipeline_convolution_pack4_3x3s1d1_winograd23_transform_output = new Pipeline(vkdev);
                 pipeline_convolution_pack4_3x3s1d1_winograd23_transform_output->set_local_size_xyz(8, 8, 1);
-                pipeline_convolution_pack4_3x3s1d1_winograd23_transform_output->create("convolution_pack4_3x3s1d1_winograd23_transform_output", specializations, 3, 7);
+                pipeline_convolution_pack4_3x3s1d1_winograd23_transform_output->create("convolution_pack4_3x3s1d1_winograd23_transform_output", opt, specializations, 3, 7);
             }
         }
     }
@@ -190,7 +190,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_convolution_pack1to4 = new Pipeline(vkdev);
         pipeline_convolution_pack1to4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_convolution_pack1to4->create("convolution_pack1to4", specializations, 4, 10);
+        pipeline_convolution_pack1to4->create("convolution_pack1to4", opt, specializations, 4, 10);
     }
 
     // pack4to1
@@ -198,7 +198,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_convolution_pack4to1 = new Pipeline(vkdev);
         pipeline_convolution_pack4to1->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_convolution_pack4to1->create("convolution_pack4to1", specializations, 4, 10);
+        pipeline_convolution_pack4to1->create("convolution_pack4to1", opt, specializations, 4, 10);
     }
 
     // fc
@@ -215,7 +215,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_innerproduct = new Pipeline(vkdev);
             pipeline_innerproduct->set_optimal_local_size_xyz(num_output, 1, 1);
-            pipeline_innerproduct->create("innerproduct", specializations, 4, 10);
+            pipeline_innerproduct->create("innerproduct", opt, specializations, 4, 10);
         }
 
         // pack4
@@ -223,7 +223,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_innerproduct_pack4 = new Pipeline(vkdev);
             pipeline_innerproduct_pack4->set_optimal_local_size_xyz(num_output / 4, 1, 1);
-            pipeline_innerproduct_pack4->create("innerproduct_pack4", specializations, 4, 10);
+            pipeline_innerproduct_pack4->create("innerproduct_pack4", opt, specializations, 4, 10);
         }
 
         // pack1to4
@@ -231,7 +231,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_innerproduct_pack1to4 = new Pipeline(vkdev);
             pipeline_innerproduct_pack1to4->set_optimal_local_size_xyz(num_output / 4, 1, 1);
-            pipeline_innerproduct_pack1to4->create("innerproduct_pack1to4", specializations, 4, 10);
+            pipeline_innerproduct_pack1to4->create("innerproduct_pack1to4", opt, specializations, 4, 10);
         }
 
         // pack4to1
@@ -239,7 +239,7 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_innerproduct_pack4to1 = new Pipeline(vkdev);
             pipeline_innerproduct_pack4to1->set_optimal_local_size_xyz(num_output, 1, 1);
-            pipeline_innerproduct_pack4to1->create("innerproduct_pack4to1", specializations, 4, 10);
+            pipeline_innerproduct_pack4to1->create("innerproduct_pack4to1", opt, specializations, 4, 10);
         }
     }
 
@@ -313,7 +313,7 @@ int Convolution_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int Convolution_vulkan::upload_model(VkTransfer& cmd)
+int Convolution_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
     const int maxk = kernel_w * kernel_h;
     int num_input = weight_data_size / maxk / num_output;
@@ -321,7 +321,7 @@ int Convolution_vulkan::upload_model(VkTransfer& cmd)
     // pack1
     if (num_input % 4 != 0 && num_output % 4 != 0)
     {
-        cmd.record_upload(weight_data, weight_data_gpu);
+        cmd.record_upload(weight_data, weight_data_gpu, opt);
     }
 
     // pack4
@@ -396,7 +396,7 @@ int Convolution_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4);
+        cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4, opt);
 
         if (kernel_w == 3 && kernel_h == 3 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1 && num_input >= 16 && num_output >= 16)
         {
@@ -514,7 +514,7 @@ int Convolution_vulkan::upload_model(VkTransfer& cmd)
                 }
             }
 
-            cmd.record_upload(weight_data_pack4_tm, weight_data_gpu_pack4_tm);
+            cmd.record_upload(weight_data_pack4_tm, weight_data_gpu_pack4_tm, opt);
         }
     }
 
@@ -560,7 +560,7 @@ int Convolution_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack1to4, weight_data_gpu_pack1to4);
+        cmd.record_upload(weight_data_pack1to4, weight_data_gpu_pack1to4, opt);
     }
 
     // pack4to1
@@ -601,21 +601,21 @@ int Convolution_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4to1, weight_data_gpu_pack4to1);
+        cmd.record_upload(weight_data_pack4to1, weight_data_gpu_pack4to1, opt);
     }
 
     if (bias_term)
     {
         if (num_output % 4 != 0)
         {
-            cmd.record_upload(bias_data, bias_data_gpu);
+            cmd.record_upload(bias_data, bias_data_gpu, opt);
         }
 
         if (num_output % 4 == 0)
         {
             Mat bias_data_pack4;
             convert_packing(bias_data, bias_data_pack4, 4);
-            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4, opt);
         }
     }
 
@@ -753,7 +753,7 @@ int Convolution_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCom
     int out_packing = num_output % 4 == 0 ? 4 : 1;
     size_t out_elemsize = elemsize / packing * out_packing;
 
-    if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_packed && !opt.use_fp16_storage)
     {
         if (out_packing == 4) out_elemsize = 4*2u;
         if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/convolution_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
 

src/layer/vulkan/convolutiondepthwise_vulkan.cpp

@@ -80,7 +80,7 @@ int ConvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_convolutiondepthwise = new Pipeline(vkdev);
             pipeline_convolutiondepthwise->set_optimal_local_size_xyz(32, 32, num_output);
-            pipeline_convolutiondepthwise->create("convolutiondepthwise", specializations, 4, 10);
+            pipeline_convolutiondepthwise->create("convolutiondepthwise", opt, specializations, 4, 10);
         }
 
         // pack4
@@ -88,7 +88,7 @@ int ConvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_convolutiondepthwise_pack4 = new Pipeline(vkdev);
             pipeline_convolutiondepthwise_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 4));
-            pipeline_convolutiondepthwise_pack4->create("convolutiondepthwise_pack4", specializations, 4, 10);
+            pipeline_convolutiondepthwise_pack4->create("convolutiondepthwise_pack4", opt, specializations, 4, 10);
         }
 
         return 0;
@@ -103,7 +103,7 @@ int ConvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_convolutiondepthwise_group = new Pipeline(vkdev);
         pipeline_convolutiondepthwise_group->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_convolutiondepthwise_group->create("convolutiondepthwise_group", specializations, 4, 10);
+        pipeline_convolutiondepthwise_group->create("convolutiondepthwise_group", opt, specializations, 4, 10);
     }
 
     // pack4
@@ -111,7 +111,7 @@ int ConvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_convolutiondepthwise_group_pack4 = new Pipeline(vkdev);
         pipeline_convolutiondepthwise_group_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_convolutiondepthwise_group_pack4->create("convolutiondepthwise_group_pack4", specializations, 4, 10);
+        pipeline_convolutiondepthwise_group_pack4->create("convolutiondepthwise_group_pack4", opt, specializations, 4, 10);
     }
 
     // pack1to4
@@ -119,7 +119,7 @@ int ConvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_convolutiondepthwise_group_pack1to4 = new Pipeline(vkdev);
         pipeline_convolutiondepthwise_group_pack1to4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_convolutiondepthwise_group_pack1to4->create("convolutiondepthwise_group_pack1to4", specializations, 4, 10);
+        pipeline_convolutiondepthwise_group_pack1to4->create("convolutiondepthwise_group_pack1to4", opt, specializations, 4, 10);
     }
 
     // pack4to1
@@ -127,7 +127,7 @@ int ConvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_convolutiondepthwise_group_pack4to1 = new Pipeline(vkdev);
         pipeline_convolutiondepthwise_group_pack4to1->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_convolutiondepthwise_group_pack4to1->create("convolutiondepthwise_group_pack4to1", specializations, 4, 10);
+        pipeline_convolutiondepthwise_group_pack4to1->create("convolutiondepthwise_group_pack4to1", opt, specializations, 4, 10);
     }
 
     if (channels % 4 == 0 && channels_g % 4 != 0)
@@ -203,7 +203,7 @@ int ConvolutionDepthWise_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int ConvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
+int ConvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
     const int maxk = kernel_w * kernel_h;
     int channels = (weight_data_size / group) / maxk / (num_output / group) * group;
@@ -214,7 +214,7 @@ int ConvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
         // pack1
         if (num_output % 4 != 0)
         {
-            cmd.record_upload(weight_data, weight_data_gpu);
+            cmd.record_upload(weight_data, weight_data_gpu, opt);
         }
 
         // pack4
@@ -224,21 +224,21 @@ int ConvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
             Mat weight_data_r2 = weight_data.reshape(maxk, group);
             convert_packing(weight_data_r2, weight_data_pack4, 4);
 
-            cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4);
+            cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4, opt);
         }
 
         if (bias_term)
         {
             if (num_output % 4 != 0)
             {
-                cmd.record_upload(bias_data, bias_data_gpu);
+                cmd.record_upload(bias_data, bias_data_gpu, opt);
             }
 
             if (num_output % 4 == 0)
             {
                 Mat bias_data_pack4;
                 convert_packing(bias_data, bias_data_pack4, 4);
-                cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+                cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4, opt);
             }
         }
 
@@ -252,7 +252,7 @@ int ConvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
     // pack1
     if (channels_g % 4 != 0 && num_output_g % 4 != 0)
     {
-        cmd.record_upload(weight_data, weight_data_gpu);
+        cmd.record_upload(weight_data, weight_data_gpu, opt);
     }
 
     // pack4
@@ -334,7 +334,7 @@ int ConvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4_groups, weight_data_gpu_pack4);
+        cmd.record_upload(weight_data_pack4_groups, weight_data_gpu_pack4, opt);
     }
 
     // pack1to4
@@ -386,7 +386,7 @@ int ConvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack1to4_groups, weight_data_gpu_pack1to4);
+        cmd.record_upload(weight_data_pack1to4_groups, weight_data_gpu_pack1to4, opt);
     }
 
     // pack4to1
@@ -434,21 +434,21 @@ int ConvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4to1_groups, weight_data_gpu_pack4to1);
+        cmd.record_upload(weight_data_pack4to1_groups, weight_data_gpu_pack4to1, opt);
     }
 
     if (bias_term)
     {
         if (num_output_g % 4 != 0)
         {
-            cmd.record_upload(bias_data, bias_data_gpu);
+            cmd.record_upload(bias_data, bias_data_gpu, opt);
         }
 
         if (num_output_g % 4 == 0)
         {
             Mat bias_data_pack4;
             convert_packing(bias_data, bias_data_pack4, 4);
-            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4, opt);
         }
     }
 
@@ -513,7 +513,7 @@ int ConvolutionDepthWise_vulkan::forward(const VkMat& bottom_blob, VkMat& top_bl
     int out_packing = num_output % 4 == 0 ? 4 : 1;
     size_t out_elemsize = elemsize / packing * out_packing;
 
-    if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_packed && !opt.use_fp16_storage)
     {
         if (out_packing == 4) out_elemsize = 4*2u;
         if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/convolutiondepthwise_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
 

src/layer/vulkan/crop_vulkan.cpp

@@ -35,14 +35,14 @@ int Crop_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_crop = new Pipeline(vkdev);
         pipeline_crop->set_optimal_local_size_xyz();
-        pipeline_crop->create("crop", specializations, 2, 13);
+        pipeline_crop->create("crop", opt, specializations, 2, 13);
     }
 
     // pack4
     {
         pipeline_crop_pack4 = new Pipeline(vkdev);
         pipeline_crop_pack4->set_optimal_local_size_xyz();
-        pipeline_crop_pack4->create("crop_pack4", specializations, 2, 13);
+        pipeline_crop_pack4->create("crop_pack4", opt, specializations, 2, 13);
     }
 
     return 0;
@@ -98,7 +98,7 @@ int Crop_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& c
     int out_packing = _outc % 4 == 0 ? 4 : 1;
     size_t out_elemsize = elemsize / packing * out_packing;
 
-    if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_packed && !opt.use_fp16_storage)
     {
         if (out_packing == 4) out_elemsize = 4*2u;
         if (out_packing == 1) out_elemsize = 4u;
@@ -195,7 +195,7 @@ int Crop_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkM
     int out_packing = _outc % 4 == 0 ? 4 : 1;
     size_t out_elemsize = elemsize / packing * out_packing;
 
-    if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_packed && !opt.use_fp16_storage)
     {
         if (out_packing == 4) out_elemsize = 4*2u;
         if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/deconvolution_vulkan.cpp

@@ -67,7 +67,7 @@ int Deconvolution_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_deconvolution = new Pipeline(vkdev);
         pipeline_deconvolution->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_deconvolution->create("deconvolution", specializations, 4, 10);
+        pipeline_deconvolution->create("deconvolution", opt, specializations, 4, 10);
     }
 
     // pack4
@@ -75,7 +75,7 @@ int Deconvolution_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_deconvolution_pack4 = new Pipeline(vkdev);
         pipeline_deconvolution_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_deconvolution_pack4->create("deconvolution_pack4", specializations, 4, 10);
+        pipeline_deconvolution_pack4->create("deconvolution_pack4", opt, specializations, 4, 10);
     }
 
     // pack1to4
@@ -83,7 +83,7 @@ int Deconvolution_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_deconvolution_pack1to4 = new Pipeline(vkdev);
         pipeline_deconvolution_pack1to4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_deconvolution_pack1to4->create("deconvolution_pack1to4", specializations, 4, 10);
+        pipeline_deconvolution_pack1to4->create("deconvolution_pack1to4", opt, specializations, 4, 10);
     }
 
     // pack4to1
@@ -91,7 +91,7 @@ int Deconvolution_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_deconvolution_pack4to1 = new Pipeline(vkdev);
         pipeline_deconvolution_pack4to1->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_deconvolution_pack4to1->create("deconvolution_pack4to1", specializations, 4, 10);
+        pipeline_deconvolution_pack4to1->create("deconvolution_pack4to1", opt, specializations, 4, 10);
     }
 
     return 0;
@@ -121,7 +121,7 @@ int Deconvolution_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int Deconvolution_vulkan::upload_model(VkTransfer& cmd)
+int Deconvolution_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
     const int maxk = kernel_w * kernel_h;
     int num_input = weight_data_size / maxk / num_output;
@@ -146,7 +146,7 @@ int Deconvolution_vulkan::upload_model(VkTransfer& cmd)
     // pack1
     if (num_input % 4 != 0 && num_output % 4 != 0)
     {
-        cmd.record_upload(weight_data_transposed, weight_data_gpu);
+        cmd.record_upload(weight_data_transposed, weight_data_gpu, opt);
     }
 
     // pack4
@@ -221,7 +221,7 @@ int Deconvolution_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4);
+        cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4, opt);
     }
 
     // pack1to4
@@ -266,7 +266,7 @@ int Deconvolution_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack1to4, weight_data_gpu_pack1to4);
+        cmd.record_upload(weight_data_pack1to4, weight_data_gpu_pack1to4, opt);
     }
 
     // pack4to1
@@ -307,21 +307,21 @@ int Deconvolution_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4to1, weight_data_gpu_pack4to1);
+        cmd.record_upload(weight_data_pack4to1, weight_data_gpu_pack4to1, opt);
     }
 
     if (bias_term)
     {
         if (num_output % 4 != 0)
         {
-            cmd.record_upload(bias_data, bias_data_gpu);
+            cmd.record_upload(bias_data, bias_data_gpu, opt);
         }
 
         if (num_output % 4 == 0)
         {
             Mat bias_data_pack4;
             convert_packing(bias_data, bias_data_pack4, 4);
-            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4, opt);
         }
     }
 

src/layer/vulkan/deconvolution_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
 

src/layer/vulkan/deconvolutiondepthwise_vulkan.cpp

@@ -77,7 +77,7 @@ int DeconvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_deconvolutiondepthwise = new Pipeline(vkdev);
             pipeline_deconvolutiondepthwise->set_optimal_local_size_xyz(32, 32, num_output);
-            pipeline_deconvolutiondepthwise->create("deconvolutiondepthwise", specializations, 4, 10);
+            pipeline_deconvolutiondepthwise->create("deconvolutiondepthwise", opt, specializations, 4, 10);
         }
 
         // pack4
@@ -85,7 +85,7 @@ int DeconvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_deconvolutiondepthwise_pack4 = new Pipeline(vkdev);
             pipeline_deconvolutiondepthwise_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 4));
-            pipeline_deconvolutiondepthwise_pack4->create("deconvolutiondepthwise_pack4", specializations, 4, 10);
+            pipeline_deconvolutiondepthwise_pack4->create("deconvolutiondepthwise_pack4", opt, specializations, 4, 10);
         }
 
         return 0;
@@ -100,7 +100,7 @@ int DeconvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_deconvolutiondepthwise_group = new Pipeline(vkdev);
         pipeline_deconvolutiondepthwise_group->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_deconvolutiondepthwise_group->create("deconvolutiondepthwise_group", specializations, 4, 10);
+        pipeline_deconvolutiondepthwise_group->create("deconvolutiondepthwise_group", opt, specializations, 4, 10);
     }
 
     // pack4
@@ -108,7 +108,7 @@ int DeconvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_deconvolutiondepthwise_group_pack4 = new Pipeline(vkdev);
         pipeline_deconvolutiondepthwise_group_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_deconvolutiondepthwise_group_pack4->create("deconvolutiondepthwise_group_pack4", specializations, 4, 10);
+        pipeline_deconvolutiondepthwise_group_pack4->create("deconvolutiondepthwise_group_pack4", opt, specializations, 4, 10);
     }
 
     // pack1to4
@@ -116,7 +116,7 @@ int DeconvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_deconvolutiondepthwise_group_pack1to4 = new Pipeline(vkdev);
         pipeline_deconvolutiondepthwise_group_pack1to4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_deconvolutiondepthwise_group_pack1to4->create("deconvolutiondepthwise_group_pack1to4", specializations, 4, 10);
+        pipeline_deconvolutiondepthwise_group_pack1to4->create("deconvolutiondepthwise_group_pack1to4", opt, specializations, 4, 10);
     }
 
     // pack4to1
@@ -124,7 +124,7 @@ int DeconvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_deconvolutiondepthwise_group_pack4to1 = new Pipeline(vkdev);
         pipeline_deconvolutiondepthwise_group_pack4to1->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
-        pipeline_deconvolutiondepthwise_group_pack4to1->create("deconvolutiondepthwise_group_pack4to1", specializations, 4, 10);
+        pipeline_deconvolutiondepthwise_group_pack4to1->create("deconvolutiondepthwise_group_pack4to1", opt, specializations, 4, 10);
     }
 
     if (channels % 4 == 0 && channels_g % 4 != 0)
@@ -200,7 +200,7 @@ int DeconvolutionDepthWise_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int DeconvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
+int DeconvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
     const int maxk = kernel_w * kernel_h;
     int channels = (weight_data_size / group) / maxk / (num_output / group) * group;
@@ -228,7 +228,7 @@ int DeconvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
         // pack1
         if (num_output % 4 != 0)
         {
-            cmd.record_upload(weight_data_transposed, weight_data_gpu);
+            cmd.record_upload(weight_data_transposed, weight_data_gpu, opt);
         }
 
         // pack4
@@ -241,21 +241,21 @@ int DeconvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
             convert_packing(weight_data_r2, weight_data_pack4, 4);
 
             weight_data_pack4 = weight_data_pack4.reshape(maxk * (group/4));
-            cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4);
+            cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4, opt);
         }
 
         if (bias_term)
         {
             if (num_output % 4 != 0)
             {
-                cmd.record_upload(bias_data, bias_data_gpu);
+                cmd.record_upload(bias_data, bias_data_gpu, opt);
             }
 
             if (num_output % 4 == 0)
             {
                 Mat bias_data_pack4;
                 convert_packing(bias_data, bias_data_pack4, 4);
-                cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+                cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4, opt);
             }
         }
 
@@ -269,7 +269,7 @@ int DeconvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
     // pack1
     if (channels_g % 4 != 0 && num_output_g % 4 != 0)
     {
-        cmd.record_upload(weight_data_transposed, weight_data_gpu);
+        cmd.record_upload(weight_data_transposed, weight_data_gpu, opt);
     }
 
     // pack4
@@ -351,7 +351,7 @@ int DeconvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4_groups, weight_data_gpu_pack4);
+        cmd.record_upload(weight_data_pack4_groups, weight_data_gpu_pack4, opt);
     }
 
     // pack1to4
@@ -403,7 +403,7 @@ int DeconvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack1to4_groups, weight_data_gpu_pack1to4);
+        cmd.record_upload(weight_data_pack1to4_groups, weight_data_gpu_pack1to4, opt);
     }
 
     // pack4to1
@@ -451,21 +451,21 @@ int DeconvolutionDepthWise_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4to1_groups, weight_data_gpu_pack4to1);
+        cmd.record_upload(weight_data_pack4to1_groups, weight_data_gpu_pack4to1, opt);
     }
 
     if (bias_term)
     {
         if (num_output_g % 4 != 0)
         {
-            cmd.record_upload(bias_data, bias_data_gpu);
+            cmd.record_upload(bias_data, bias_data_gpu, opt);
         }
 
         if (num_output_g % 4 == 0)
         {
             Mat bias_data_pack4;
             convert_packing(bias_data, bias_data_pack4, 4);
-            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4, opt);
         }
     }
 

src/layer/vulkan/deconvolutiondepthwise_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
 

src/layer/vulkan/dropout_vulkan.cpp

@@ -29,19 +29,21 @@ Dropout_vulkan::Dropout_vulkan()
 
 int Dropout_vulkan::create_pipeline(const Option& opt)
 {
-    pipeline_dropout = new Pipeline(vkdev);
-    pipeline_dropout->set_optimal_local_size_xyz();
-
     std::vector<vk_specialization_type> specializations(1);
     specializations[0].f = scale;
 
-    pipeline_dropout->create("dropout", specializations, 1, 5);
+    // pack1
+    {
+        pipeline_dropout = new Pipeline(vkdev);
+        pipeline_dropout->set_optimal_local_size_xyz();
+        pipeline_dropout->create("dropout", opt, specializations, 1, 5);
+    }
 
     // pack4
     {
         pipeline_dropout_pack4 = new Pipeline(vkdev);
         pipeline_dropout_pack4->set_optimal_local_size_xyz();
-        pipeline_dropout_pack4->create("dropout_pack4", specializations, 1, 5);
+        pipeline_dropout_pack4->create("dropout_pack4", opt, specializations, 1, 5);
     }
 
     return 0;

src/layer/vulkan/eltwise_vulkan.cpp

@@ -39,20 +39,20 @@ int Eltwise_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_eltwise[0] = new Pipeline(vkdev);
         pipeline_eltwise[0]->set_optimal_local_size_xyz();
-        pipeline_eltwise[0]->create("eltwise", specializations, 3, 5+2);
+        pipeline_eltwise[0]->create("eltwise", opt, specializations, 3, 5+2);
         pipeline_eltwise[1] = new Pipeline(vkdev);
         pipeline_eltwise[1]->set_optimal_local_size_xyz();
-        pipeline_eltwise[1]->create("eltwise", specializations, 3, 5+2);
+        pipeline_eltwise[1]->create("eltwise", opt, specializations, 3, 5+2);
     }
 
     // pack4
     {
         pipeline_eltwise_pack4[0] = new Pipeline(vkdev);
         pipeline_eltwise_pack4[0]->set_optimal_local_size_xyz();
-        pipeline_eltwise_pack4[0]->create("eltwise_pack4", specializations, 3, 5+2);
+        pipeline_eltwise_pack4[0]->create("eltwise_pack4", opt, specializations, 3, 5+2);
         pipeline_eltwise_pack4[1] = new Pipeline(vkdev);
         pipeline_eltwise_pack4[1]->set_optimal_local_size_xyz();
-        pipeline_eltwise_pack4[1]->create("eltwise_pack4", specializations, 3, 5+2);
+        pipeline_eltwise_pack4[1]->create("eltwise_pack4", opt, specializations, 3, 5+2);
     }
 
     return 0;

src/layer/vulkan/flatten_vulkan.cpp

@@ -35,21 +35,21 @@ int Flatten_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_flatten = new Pipeline(vkdev);
         pipeline_flatten->set_optimal_local_size_xyz();
-        pipeline_flatten->create("flatten", specializations, 2, 10);
+        pipeline_flatten->create("flatten", opt, specializations, 2, 10);
     }
 
     // pack4
     {
         pipeline_flatten_pack4 = new Pipeline(vkdev);
         pipeline_flatten_pack4->set_optimal_local_size_xyz();
-        pipeline_flatten_pack4->create("flatten_pack4", specializations, 2, 10);
+        pipeline_flatten_pack4->create("flatten_pack4", opt, specializations, 2, 10);
     }
 
     // pack1to4
     {
         pipeline_flatten_pack1to4 = new Pipeline(vkdev);
         pipeline_flatten_pack1to4->set_optimal_local_size_xyz();
-        pipeline_flatten_pack1to4->create("flatten_pack1to4", specializations, 2, 10);
+        pipeline_flatten_pack1to4->create("flatten_pack1to4", opt, specializations, 2, 10);
     }
 
     return 0;
@@ -90,7 +90,7 @@ int Flatten_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
     int out_packing = total % 4 == 0 ? 4 : 1;
     size_t out_elemsize = elemsize / packing * out_packing;
 
-    if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_packed && !opt.use_fp16_storage)
     {
         if (out_packing == 4) out_elemsize = 4*2u;
         if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/innerproduct_vulkan.cpp

@@ -59,7 +59,7 @@ int InnerProduct_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_innerproduct = new Pipeline(vkdev);
         pipeline_innerproduct->set_optimal_local_size_xyz(num_output, 1, 1);
-        pipeline_innerproduct->create("innerproduct", specializations, 4, 10);
+        pipeline_innerproduct->create("innerproduct", opt, specializations, 4, 10);
     }
 
     // pack4
@@ -67,12 +67,12 @@ int InnerProduct_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_innerproduct_pack4 = new Pipeline(vkdev);
         pipeline_innerproduct_pack4->set_optimal_local_size_xyz(num_output / 4, 1, 1);
-        pipeline_innerproduct_pack4->create("innerproduct_pack4", specializations, 4, 10);
+        pipeline_innerproduct_pack4->create("innerproduct_pack4", opt, specializations, 4, 10);
 
         {
             pipeline_innerproduct_pack4_lds_64 = new Pipeline(vkdev);
             pipeline_innerproduct_pack4_lds_64->set_local_size_xyz(64, 1, 1);
-            pipeline_innerproduct_pack4_lds_64->create("innerproduct_pack4_lds_64", specializations, 4, 10);
+            pipeline_innerproduct_pack4_lds_64->create("innerproduct_pack4_lds_64", opt, specializations, 4, 10);
         }
     }
 
@@ -81,7 +81,7 @@ int InnerProduct_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_innerproduct_pack1to4 = new Pipeline(vkdev);
         pipeline_innerproduct_pack1to4->set_optimal_local_size_xyz(num_output / 4, 1, 1);
-        pipeline_innerproduct_pack1to4->create("innerproduct_pack1to4", specializations, 4, 10);
+        pipeline_innerproduct_pack1to4->create("innerproduct_pack1to4", opt, specializations, 4, 10);
     }
 
     // pack4to1
@@ -89,7 +89,7 @@ int InnerProduct_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_innerproduct_pack4to1 = new Pipeline(vkdev);
         pipeline_innerproduct_pack4to1->set_optimal_local_size_xyz(num_output, 1, 1);
-        pipeline_innerproduct_pack4to1->create("innerproduct_pack4to1", specializations, 4, 10);
+        pipeline_innerproduct_pack4to1->create("innerproduct_pack4to1", opt, specializations, 4, 10);
     }
 
     return 0;
@@ -122,14 +122,14 @@ int InnerProduct_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int InnerProduct_vulkan::upload_model(VkTransfer& cmd)
+int InnerProduct_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
     int num_input = weight_data_size / num_output;
 
     // pack1
     if (num_input % 4 != 0 && num_output % 4 != 0)
     {
-        cmd.record_upload(weight_data, weight_data_gpu);
+        cmd.record_upload(weight_data, weight_data_gpu, opt);
     }
 
     // pack4
@@ -183,7 +183,7 @@ int InnerProduct_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4);
+        cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4, opt);
     }
 
     // pack1to4
@@ -218,7 +218,7 @@ int InnerProduct_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack1to4, weight_data_gpu_pack1to4);
+        cmd.record_upload(weight_data_pack1to4, weight_data_gpu_pack1to4, opt);
     }
 
     // pack4to1
@@ -251,21 +251,21 @@ int InnerProduct_vulkan::upload_model(VkTransfer& cmd)
             }
         }
 
-        cmd.record_upload(weight_data_pack4to1, weight_data_gpu_pack4to1);
+        cmd.record_upload(weight_data_pack4to1, weight_data_gpu_pack4to1, opt);
     }
 
     if (bias_term)
     {
         if (num_output % 4 != 0)
         {
-            cmd.record_upload(bias_data, bias_data_gpu);
+            cmd.record_upload(bias_data, bias_data_gpu, opt);
         }
 
         if (num_output % 4 == 0)
         {
             Mat bias_data_pack4;
             convert_packing(bias_data, bias_data_pack4, 4);
-            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4, opt);
         }
     }
 
@@ -290,7 +290,7 @@ int InnerProduct_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCo
     int out_packing = num_output % 4 == 0 ? 4 : 1;
     size_t out_elemsize = elemsize / packing * out_packing;
 
-    if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_packed && !opt.use_fp16_storage)
     {
         if (out_packing == 4) out_elemsize = 4*2u;
         if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/innerproduct_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
 

src/layer/vulkan/interp_vulkan.cpp

@@ -43,14 +43,14 @@ int Interp_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_interp = new Pipeline(vkdev);
             pipeline_interp->set_optimal_local_size_xyz();
-            pipeline_interp->create("interp", specializations, 2, 12);
+            pipeline_interp->create("interp", opt, specializations, 2, 12);
         }
 
         // pack4
         {
             pipeline_interp_pack4 = new Pipeline(vkdev);
             pipeline_interp_pack4->set_optimal_local_size_xyz();
-            pipeline_interp_pack4->create("interp_pack4", specializations, 2, 12);
+            pipeline_interp_pack4->create("interp_pack4", opt, specializations, 2, 12);
         }
     }
 
@@ -60,24 +60,24 @@ int Interp_vulkan::create_pipeline(const Option& opt)
 
         pipeline_interp_bicubic_coeffs_x = new Pipeline(vkdev);
         pipeline_interp_bicubic_coeffs_x->set_optimal_local_size_xyz(64, 1, 1);
-        pipeline_interp_bicubic_coeffs_x->create("interp_bicubic_coeffs", specializations, 2, 3);
+        pipeline_interp_bicubic_coeffs_x->create("interp_bicubic_coeffs", opt, specializations, 2, 3);
 
         pipeline_interp_bicubic_coeffs_y = new Pipeline(vkdev);
         pipeline_interp_bicubic_coeffs_y->set_optimal_local_size_xyz(64, 1, 1);
-        pipeline_interp_bicubic_coeffs_y->create("interp_bicubic_coeffs", specializations, 2, 3);
+        pipeline_interp_bicubic_coeffs_y->create("interp_bicubic_coeffs", opt, specializations, 2, 3);
 
         // pack1
         {
             pipeline_interp_bicubic = new Pipeline(vkdev);
             pipeline_interp_bicubic->set_optimal_local_size_xyz();
-            pipeline_interp_bicubic->create("interp_bicubic", specializations, 6, 10);
+            pipeline_interp_bicubic->create("interp_bicubic", opt, specializations, 6, 10);
         }
 
         // pack4
         {
             pipeline_interp_bicubic_pack4 = new Pipeline(vkdev);
             pipeline_interp_bicubic_pack4->set_optimal_local_size_xyz();
-            pipeline_interp_bicubic_pack4->create("interp_bicubic_pack4", specializations, 6, 10);
+            pipeline_interp_bicubic_pack4->create("interp_bicubic_pack4", opt, specializations, 6, 10);
         }
     }
 

src/layer/vulkan/lrn_vulkan.cpp

@@ -52,20 +52,20 @@ int LRN_vulkan::create_pipeline(const Option& opt)
             specializations[2].i = local_size - pad - 1;
         }
 
-        pipeline_lrn_square_pad->create("lrn_square_pad", specializations, 2, 10);
+        pipeline_lrn_square_pad->create("lrn_square_pad", opt, specializations, 2, 10);
 
         // pack4
         if (region_type == 0)
         {
             pipeline_lrn_square_pad_across_channel_pack4 = new Pipeline(vkdev);
             pipeline_lrn_square_pad_across_channel_pack4->set_optimal_local_size_xyz();
-            pipeline_lrn_square_pad_across_channel_pack4->create("lrn_square_pad_across_channel_pack4", specializations, 2, 10);
+            pipeline_lrn_square_pad_across_channel_pack4->create("lrn_square_pad_across_channel_pack4", opt, specializations, 2, 10);
         }
         if (region_type == 1)
         {
             pipeline_lrn_square_pad_within_channel_pack4 = new Pipeline(vkdev);
             pipeline_lrn_square_pad_within_channel_pack4->set_optimal_local_size_xyz();
-            pipeline_lrn_square_pad_within_channel_pack4->create("lrn_square_pad_within_channel_pack4", specializations, 2, 10);
+            pipeline_lrn_square_pad_within_channel_pack4->create("lrn_square_pad_within_channel_pack4", opt, specializations, 2, 10);
         }
     }
 
@@ -80,20 +80,20 @@ int LRN_vulkan::create_pipeline(const Option& opt)
         specializations[3].f = beta;
         specializations[4].f = bias;
 
-        pipeline_lrn_norm->create("lrn_norm", specializations, 2, 10);
+        pipeline_lrn_norm->create("lrn_norm", opt, specializations, 2, 10);
 
         // pack4
         if (region_type == 0)
         {
             pipeline_lrn_norm_across_channel_pack4 = new Pipeline(vkdev);
             pipeline_lrn_norm_across_channel_pack4->set_optimal_local_size_xyz();
-            pipeline_lrn_norm_across_channel_pack4->create("lrn_norm_across_channel_pack4", specializations, 2, 10);
+            pipeline_lrn_norm_across_channel_pack4->create("lrn_norm_across_channel_pack4", opt, specializations, 2, 10);
         }
         if (region_type == 1)
         {
             pipeline_lrn_norm_within_channel_pack4 = new Pipeline(vkdev);
             pipeline_lrn_norm_within_channel_pack4->set_optimal_local_size_xyz();
-            pipeline_lrn_norm_within_channel_pack4->create("lrn_norm_within_channel_pack4", specializations, 2, 10);
+            pipeline_lrn_norm_within_channel_pack4->create("lrn_norm_within_channel_pack4", opt, specializations, 2, 10);
         }
     }
 

src/layer/vulkan/packing_vulkan.cpp

@@ -34,14 +34,14 @@ int Packing_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_packing_1to4 = new Pipeline(vkdev);
         pipeline_packing_1to4->set_optimal_local_size_xyz();
-        pipeline_packing_1to4->create("packing_1to4", specializations, 2, 10);
+        pipeline_packing_1to4->create("packing_1to4", opt, specializations, 2, 10);
     }
 
     if (out_packing == 1)
     {
         pipeline_packing_4to1 = new Pipeline(vkdev);
         pipeline_packing_4to1->set_optimal_local_size_xyz();
-        pipeline_packing_4to1->create("packing_4to1", specializations, 2, 10);
+        pipeline_packing_4to1->create("packing_4to1", opt, specializations, 2, 10);
     }
 
     return 0;
@@ -96,7 +96,7 @@ int Packing_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
 
     if (dims == 1)
     {
-        if (vkdev->info.support_fp16_storage && out_packing == 1)
+        if (opt.use_fp16_storage && out_packing == 1)
         {
             top_blob = bottom_blob;
             top_blob.w = w * packing;
@@ -108,7 +108,7 @@ int Packing_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
 
         int outw = (w * packing + out_packing - 1) / out_packing;
         size_t out_elemsize = elemsize / packing * out_packing;
-        if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+        if (opt.use_fp16_packed && !opt.use_fp16_storage)
         {
             if (out_packing == 4) out_elemsize = 4*2u;
             if (out_packing == 1) out_elemsize = 4u;
@@ -123,7 +123,7 @@ int Packing_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
     {
         int outh = (h * packing + out_packing - 1) / out_packing;
         size_t out_elemsize = elemsize / packing * out_packing;
-        if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+        if (opt.use_fp16_packed && !opt.use_fp16_storage)
         {
             if (out_packing == 4) out_elemsize = 4*2u;
             if (out_packing == 1) out_elemsize = 4u;
@@ -138,7 +138,7 @@ int Packing_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
     {
         int outc = (channels * packing + out_packing - 1) / out_packing;
         size_t out_elemsize = elemsize / packing * out_packing;
-        if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+        if (opt.use_fp16_packed && !opt.use_fp16_storage)
         {
             if (out_packing == 4) out_elemsize = 4*2u;
             if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/padding_vulkan.cpp

@@ -36,14 +36,14 @@ int Padding_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_padding = new Pipeline(vkdev);
         pipeline_padding->set_optimal_local_size_xyz();
-        pipeline_padding->create("padding", specializations, 2, 12);
+        pipeline_padding->create("padding", opt, specializations, 2, 12);
     }
 
     // pack4
     {
         pipeline_padding_pack4 = new Pipeline(vkdev);
         pipeline_padding_pack4->set_optimal_local_size_xyz();
-        pipeline_padding_pack4->create("padding_pack4", specializations, 2, 12);
+        pipeline_padding_pack4->create("padding_pack4", opt, specializations, 2, 12);
     }
 
     return 0;

src/layer/vulkan/permute_vulkan.cpp

@@ -35,14 +35,14 @@ int Permute_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_permute = new Pipeline(vkdev);
         pipeline_permute->set_optimal_local_size_xyz();
-        pipeline_permute->create("permute", specializations, 2, 10);
+        pipeline_permute->create("permute", opt, specializations, 2, 10);
     }
 
     // pack4
     {
         pipeline_permute_pack4to1 = new Pipeline(vkdev);
         pipeline_permute_pack4to1->set_optimal_local_size_xyz();
-        pipeline_permute_pack4to1->create("permute_pack4to1", specializations, 2, 10);
+        pipeline_permute_pack4to1->create("permute_pack4to1", opt, specializations, 2, 10);
     }
 
     return 0;
@@ -72,7 +72,7 @@ int Permute_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
     int out_packing = 1;
     size_t out_elemsize = elemsize / packing;
 
-    if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_packed && !opt.use_fp16_storage)
     {
         if (out_packing == 4) out_elemsize = 4*2u;
         if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/pooling_vulkan.cpp

@@ -72,14 +72,14 @@ int Pooling_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_pooling = new Pipeline(vkdev);
         pipeline_pooling->set_optimal_local_size_xyz();
-        pipeline_pooling->create("pooling", specializations, 2, 10);
+        pipeline_pooling->create("pooling", opt, specializations, 2, 10);
     }
 
     // pack4
     {
         pipeline_pooling_pack4 = new Pipeline(vkdev);
         pipeline_pooling_pack4->set_optimal_local_size_xyz();
-        pipeline_pooling_pack4->create("pooling_pack4", specializations, 2, 10);
+        pipeline_pooling_pack4->create("pooling_pack4", opt, specializations, 2, 10);
     }
 
     if (global_pooling)
@@ -91,14 +91,14 @@ int Pooling_vulkan::create_pipeline(const Option& opt)
         {
             pipeline_pooling_global = new Pipeline(vkdev);
             pipeline_pooling_global->set_optimal_local_size_xyz(256, 1, 1);
-            pipeline_pooling_global->create("pooling_global", specializations, 2, 10);
+            pipeline_pooling_global->create("pooling_global", opt, specializations, 2, 10);
         }
 
         // pack4
         {
             pipeline_pooling_global_pack4 = new Pipeline(vkdev);
             pipeline_pooling_global_pack4->set_optimal_local_size_xyz(256, 1, 1);
-            pipeline_pooling_global_pack4->create("pooling_global_pack4", specializations, 2, 10);
+            pipeline_pooling_global_pack4->create("pooling_global_pack4", opt, specializations, 2, 10);
         }
     }
 

src/layer/vulkan/prelu_vulkan.cpp

@@ -36,7 +36,7 @@ int PReLU_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_prelu = new Pipeline(vkdev);
         pipeline_prelu->set_optimal_local_size_xyz(8, 8, num_slope);
-        pipeline_prelu->create("prelu", specializations, 2, 5);
+        pipeline_prelu->create("prelu", opt, specializations, 2, 5);
     }
 
     // pack4
@@ -44,7 +44,7 @@ int PReLU_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_prelu_pack4 = new Pipeline(vkdev);
         pipeline_prelu_pack4->set_optimal_local_size_xyz(8, 8, num_slope / 4);
-        pipeline_prelu_pack4->create("prelu_pack4", specializations, 2, 5);
+        pipeline_prelu_pack4->create("prelu_pack4", opt, specializations, 2, 5);
     }
 
     return 0;
@@ -61,18 +61,18 @@ int PReLU_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int PReLU_vulkan::upload_model(VkTransfer& cmd)
+int PReLU_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
     if (num_slope == 1)
     {
         // dup4 for pack4
         Mat slope_data4(4);
         slope_data4.fill(slope_data[0]);
-        cmd.record_upload(slope_data4, slope_data_gpu);
+        cmd.record_upload(slope_data4, slope_data_gpu, opt);
     }
     else
     {
-        cmd.record_upload(slope_data, slope_data_gpu);
+        cmd.record_upload(slope_data, slope_data_gpu, opt);
     }
 
     return 0;

src/layer/vulkan/prelu_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
 

src/layer/vulkan/priorbox_vulkan.cpp

@@ -55,7 +55,7 @@ int PriorBox_vulkan::create_pipeline(const Option& opt)
 
         pipeline_priorbox = new Pipeline(vkdev);
         pipeline_priorbox->set_optimal_local_size_xyz();
-        pipeline_priorbox->create("priorbox", specializations, 4, 6);
+        pipeline_priorbox->create("priorbox", opt, specializations, 4, 6);
     }
 
     // mxnet style
@@ -74,7 +74,7 @@ int PriorBox_vulkan::create_pipeline(const Option& opt)
 
         pipeline_priorbox_mxnet = new Pipeline(vkdev);
         pipeline_priorbox_mxnet->set_optimal_local_size_xyz();
-        pipeline_priorbox_mxnet->create("priorbox_mxnet", specializations, 3, 4);
+        pipeline_priorbox_mxnet->create("priorbox_mxnet", opt, specializations, 3, 4);
     }
 
     return 0;
@@ -91,14 +91,14 @@ int PriorBox_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int PriorBox_vulkan::upload_model(VkTransfer& cmd)
+int PriorBox_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
-    cmd.record_upload(min_sizes, min_sizes_gpu);
+    cmd.record_upload(min_sizes, min_sizes_gpu, opt);
 
     if (max_sizes.w > 0)
-        cmd.record_upload(max_sizes, max_sizes_gpu);
+        cmd.record_upload(max_sizes, max_sizes_gpu, opt);
 
-    cmd.record_upload(aspect_ratios, aspect_ratios_gpu);
+    cmd.record_upload(aspect_ratios, aspect_ratios_gpu, opt);
 
     return 0;
 }
@@ -109,7 +109,7 @@ int PriorBox_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector
     int h = bottom_blobs[0].h;
     size_t elemsize = 4u;
 
-    if (vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_storage)
     {
         elemsize = 2u;
     }

src/layer/vulkan/priorbox_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
 

src/layer/vulkan/relu_vulkan.cpp

@@ -35,14 +35,14 @@ int ReLU_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_relu = new Pipeline(vkdev);
         pipeline_relu->set_optimal_local_size_xyz();
-        pipeline_relu->create("relu", specializations, 1, 5);
+        pipeline_relu->create("relu", opt, specializations, 1, 5);
     }
 
     // pack4
     {
         pipeline_relu_pack4 = new Pipeline(vkdev);
         pipeline_relu_pack4->set_optimal_local_size_xyz();
-        pipeline_relu_pack4->create("relu_pack4", specializations, 1, 5);
+        pipeline_relu_pack4->create("relu_pack4", opt, specializations, 1, 5);
     }
 
     return 0;

src/layer/vulkan/reorg_vulkan.cpp

@@ -36,21 +36,21 @@ int Reorg_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_reorg = new Pipeline(vkdev);
         pipeline_reorg->set_optimal_local_size_xyz();
-        pipeline_reorg->create("reorg", specializations, 2, 10);
+        pipeline_reorg->create("reorg", opt, specializations, 2, 10);
     }
 
     // pack4
     {
         pipeline_reorg_pack4 = new Pipeline(vkdev);
         pipeline_reorg_pack4->set_optimal_local_size_xyz();
-        pipeline_reorg_pack4->create("reorg_pack4", specializations, 2, 10);
+        pipeline_reorg_pack4->create("reorg_pack4", opt, specializations, 2, 10);
     }
 
     // pack1to4
     {
         pipeline_reorg_pack1to4 = new Pipeline(vkdev);
         pipeline_reorg_pack1to4->set_optimal_local_size_xyz();
-        pipeline_reorg_pack1to4->create("reorg_pack1to4", specializations, 2, 10);
+        pipeline_reorg_pack1to4->create("reorg_pack1to4", opt, specializations, 2, 10);
     }
 
     return 0;
@@ -85,7 +85,7 @@ int Reorg_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute&
     int out_packing = outc % 4 == 0 ? 4 : 1;
     size_t out_elemsize = elemsize / packing * out_packing;
 
-    if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_packed && !opt.use_fp16_storage)
     {
         if (out_packing == 4) out_elemsize = 4*2u;
         if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/reshape_vulkan.cpp

@@ -37,28 +37,28 @@ int Reshape_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_reshape = new Pipeline(vkdev);
         pipeline_reshape->set_optimal_local_size_xyz();
-        pipeline_reshape->create("reshape", specializations, 2, 10);
+        pipeline_reshape->create("reshape", opt, specializations, 2, 10);
     }
 
     // pack4
     {
         pipeline_reshape_pack4 = new Pipeline(vkdev);
         pipeline_reshape_pack4->set_optimal_local_size_xyz();
-        pipeline_reshape_pack4->create("reshape_pack4", specializations, 2, 10);
+        pipeline_reshape_pack4->create("reshape_pack4", opt, specializations, 2, 10);
     }
 
     // pack1to4
     {
         pipeline_reshape_pack1to4 = new Pipeline(vkdev);
         pipeline_reshape_pack1to4->set_optimal_local_size_xyz();
-        pipeline_reshape_pack1to4->create("reshape_pack1to4", specializations, 2, 10);
+        pipeline_reshape_pack1to4->create("reshape_pack1to4", opt, specializations, 2, 10);
     }
 
     // pack4to1
     {
         pipeline_reshape_pack4to1 = new Pipeline(vkdev);
         pipeline_reshape_pack4to1->set_optimal_local_size_xyz();
-        pipeline_reshape_pack4to1->create("reshape_pack4to1", specializations, 2, 10);
+        pipeline_reshape_pack4to1->create("reshape_pack4to1", opt, specializations, 2, 10);
     }
 
     return 0;
@@ -105,7 +105,7 @@ int Reshape_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
         out_packing = _w % 4 == 0 ? 4 : 1;
         size_t out_elemsize = elemsize / packing * out_packing;
 
-        if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+        if (opt.use_fp16_packed && !opt.use_fp16_storage)
         {
             if (out_packing == 4) out_elemsize = 4*2u;
             if (out_packing == 1) out_elemsize = 4u;
@@ -137,7 +137,7 @@ int Reshape_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
         out_packing = _h % 4 == 0 ? 4 : 1;
         size_t out_elemsize = elemsize / packing * out_packing;
 
-        if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+        if (opt.use_fp16_packed && !opt.use_fp16_storage)
         {
             if (out_packing == 4) out_elemsize = 4*2u;
             if (out_packing == 1) out_elemsize = 4u;
@@ -174,7 +174,7 @@ int Reshape_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
         out_packing = _c % 4 == 0 ? 4 : 1;
         size_t out_elemsize = elemsize / packing * out_packing;
 
-        if (vkdev->info.support_fp16_packed && !vkdev->info.support_fp16_storage)
+        if (opt.use_fp16_packed && !opt.use_fp16_storage)
         {
             if (out_packing == 4) out_elemsize = 4*2u;
             if (out_packing == 1) out_elemsize = 4u;

src/layer/vulkan/scale_vulkan.cpp

@@ -35,13 +35,13 @@ int Scale_vulkan::create_pipeline(const Option& opt)
 
         pipeline_scale = new Pipeline(vkdev);
         pipeline_scale->set_optimal_local_size_xyz();
-        pipeline_scale->create("scale", specializations, 3, 5);
+        pipeline_scale->create("scale", opt, specializations, 3, 5);
 
         // pack4
         {
             pipeline_scale_pack4 = new Pipeline(vkdev);
             pipeline_scale_pack4->set_optimal_local_size_xyz();
-            pipeline_scale_pack4->create("scale_pack4", specializations, 3, 5);
+            pipeline_scale_pack4->create("scale_pack4", opt, specializations, 3, 5);
         }
 
         return 0;
@@ -55,7 +55,7 @@ int Scale_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_scale = new Pipeline(vkdev);
         pipeline_scale->set_optimal_local_size_xyz(8, 8, scale_data_size);
-        pipeline_scale->create("scale", specializations, 3, 5);
+        pipeline_scale->create("scale", opt, specializations, 3, 5);
     }
 
     // pack4
@@ -63,7 +63,7 @@ int Scale_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_scale_pack4 = new Pipeline(vkdev);
         pipeline_scale_pack4->set_optimal_local_size_xyz(8, 8, scale_data_size / 4);
-        pipeline_scale_pack4->create("scale_pack4", specializations, 3, 5);
+        pipeline_scale_pack4->create("scale_pack4", opt, specializations, 3, 5);
     }
 
     return 0;
@@ -80,7 +80,7 @@ int Scale_vulkan::destroy_pipeline(const Option& opt)
     return 0;
 }
 
-int Scale_vulkan::upload_model(VkTransfer& cmd)
+int Scale_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
     if (scale_data_size == -233)
         return 0;
@@ -88,7 +88,7 @@ int Scale_vulkan::upload_model(VkTransfer& cmd)
     // pack1
     if (scale_data_size % 4 != 0)
     {
-        cmd.record_upload(scale_data, scale_data_gpu);
+        cmd.record_upload(scale_data, scale_data_gpu, opt);
     }
 
     // pack4
@@ -96,7 +96,7 @@ int Scale_vulkan::upload_model(VkTransfer& cmd)
     {
         Mat scale_data_pack4;
         convert_packing(scale_data, scale_data_pack4, 4);
-        cmd.record_upload(scale_data_pack4, scale_data_gpu_pack4);
+        cmd.record_upload(scale_data_pack4, scale_data_gpu_pack4, opt);
     }
 
     if (bias_term)
@@ -104,7 +104,7 @@ int Scale_vulkan::upload_model(VkTransfer& cmd)
         // pack1
         if (scale_data_size % 4 != 0)
         {
-            cmd.record_upload(bias_data, bias_data_gpu);
+            cmd.record_upload(bias_data, bias_data_gpu, opt);
         }
 
         // pack4
@@ -112,7 +112,7 @@ int Scale_vulkan::upload_model(VkTransfer& cmd)
         {
             Mat bias_data_pack4;
             convert_packing(bias_data, bias_data_pack4, 4);
-            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+            cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4, opt);
         }
     }
 

src/layer/vulkan/scale_vulkan.h

@@ -27,7 +27,7 @@ public:
     virtual int create_pipeline(const Option& opt);
     virtual int destroy_pipeline(const Option& opt);
 
-    virtual int upload_model(VkTransfer& cmd);
+    virtual int upload_model(VkTransfer& cmd, const Option& opt);
 
     virtual int forward_inplace(std::vector<VkMat>& bottom_top_blobs, VkCompute& cmd, const Option& opt) const;
     virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;

src/layer/vulkan/shufflechannel_vulkan.cpp

@@ -35,14 +35,14 @@ int ShuffleChannel_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_shufflechannel = new Pipeline(vkdev);
         pipeline_shufflechannel->set_optimal_local_size_xyz();
-        pipeline_shufflechannel->create("shufflechannel", specializations, 2, 10);
+        pipeline_shufflechannel->create("shufflechannel", opt, specializations, 2, 10);
     }
 
     // pack4
     {
         pipeline_shufflechannel_pack4 = new Pipeline(vkdev);
         pipeline_shufflechannel_pack4->set_optimal_local_size_xyz();
-        pipeline_shufflechannel_pack4->create("shufflechannel_pack4", specializations, 2, 10);
+        pipeline_shufflechannel_pack4->create("shufflechannel_pack4", opt, specializations, 2, 10);
     }
 
     return 0;

src/layer/vulkan/sigmoid_vulkan.cpp

@@ -35,14 +35,14 @@ int Sigmoid_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_sigmoid = new Pipeline(vkdev);
         pipeline_sigmoid->set_optimal_local_size_xyz();
-        pipeline_sigmoid->create("sigmoid", specializations, 1, 5);
+        pipeline_sigmoid->create("sigmoid", opt, specializations, 1, 5);
     }
 
     // pack4
     {
         pipeline_sigmoid_pack4 = new Pipeline(vkdev);
         pipeline_sigmoid_pack4->set_optimal_local_size_xyz();
-        pipeline_sigmoid_pack4->create("sigmoid_pack4", specializations, 1, 5);
+        pipeline_sigmoid_pack4->create("sigmoid_pack4", opt, specializations, 1, 5);
     }
 
     return 0;

src/layer/vulkan/softmax_vulkan.cpp

@@ -53,10 +53,10 @@ int Softmax_vulkan::create_pipeline(const Option& opt)
         pipeline_softmax_reduce_sum->set_optimal_local_size_xyz();
         pipeline_softmax_div_sum->set_optimal_local_size_xyz();
 
-        pipeline_softmax_reduce_max->create("softmax_reduce_max", specializations, 2, 10);
-        pipeline_softmax_exp_sub_max->create("softmax_exp_sub_max", specializations, 2, 10);
-        pipeline_softmax_reduce_sum->create("softmax_reduce_sum", specializations, 2, 10);
-        pipeline_softmax_div_sum->create("softmax_div_sum", specializations, 2, 10);
+        pipeline_softmax_reduce_max->create("softmax_reduce_max", opt, specializations, 2, 10);
+        pipeline_softmax_exp_sub_max->create("softmax_exp_sub_max", opt, specializations, 2, 10);
+        pipeline_softmax_reduce_sum->create("softmax_reduce_sum", opt, specializations, 2, 10);
+        pipeline_softmax_div_sum->create("softmax_div_sum", opt, specializations, 2, 10);
     }
 
     // pack4
@@ -71,10 +71,10 @@ int Softmax_vulkan::create_pipeline(const Option& opt)
         pipeline_softmax_reduce_sum_pack4->set_optimal_local_size_xyz();
         pipeline_softmax_div_sum_pack4->set_optimal_local_size_xyz();
 
-        pipeline_softmax_reduce_max_pack4->create("softmax_reduce_max_pack4", specializations, 2, 10);
-        pipeline_softmax_exp_sub_max_pack4->create("softmax_exp_sub_max_pack4", specializations, 2, 10);
-        pipeline_softmax_reduce_sum_pack4->create("softmax_reduce_sum_pack4", specializations, 2, 10);
-        pipeline_softmax_div_sum_pack4->create("softmax_div_sum_pack4", specializations, 2, 10);
+        pipeline_softmax_reduce_max_pack4->create("softmax_reduce_max_pack4", opt, specializations, 2, 10);
+        pipeline_softmax_exp_sub_max_pack4->create("softmax_exp_sub_max_pack4", opt, specializations, 2, 10);
+        pipeline_softmax_reduce_sum_pack4->create("softmax_reduce_sum_pack4", opt, specializations, 2, 10);
+        pipeline_softmax_div_sum_pack4->create("softmax_div_sum_pack4", opt, specializations, 2, 10);
     }
 
     return 0;

src/layer/vulkan/tanh_vulkan.cpp

@@ -35,14 +35,14 @@ int TanH_vulkan::create_pipeline(const Option& opt)
     {
         pipeline_tanh = new Pipeline(vkdev);
         pipeline_tanh->set_optimal_local_size_xyz();
-        pipeline_tanh->create("tanh", specializations, 1, 5);
+        pipeline_tanh->create("tanh", opt, specializations, 1, 5);
     }
 
     // pack4
     {
         pipeline_tanh_pack4 = new Pipeline(vkdev);
         pipeline_tanh_pack4->set_optimal_local_size_xyz();
-        pipeline_tanh_pack4->create("tanh_pack4", specializations, 1, 5);
+        pipeline_tanh_pack4->create("tanh_pack4", opt, specializations, 1, 5);
     }
 
     return 0;

src/layer/vulkan/unaryop_vulkan.cpp

@@ -28,19 +28,21 @@ UnaryOp_vulkan::UnaryOp_vulkan()
 
 int UnaryOp_vulkan::create_pipeline(const Option& opt)
 {
-    pipeline_unaryop = new Pipeline(vkdev);
-    pipeline_unaryop->set_optimal_local_size_xyz();
-
     std::vector<vk_specialization_type> specializations(1);
     specializations[0].i = op_type;
 
-    pipeline_unaryop->create("unaryop", specializations, 1, 5);
+    // pack1
+    {
+        pipeline_unaryop = new Pipeline(vkdev);
+        pipeline_unaryop->set_optimal_local_size_xyz();
+        pipeline_unaryop->create("unaryop", opt, specializations, 1, 5);
+    }
 
     // pack4
     {
         pipeline_unaryop_pack4 = new Pipeline(vkdev);
         pipeline_unaryop_pack4->set_optimal_local_size_xyz();
-        pipeline_unaryop_pack4->create("unaryop_pack4", specializations, 1, 5);
+        pipeline_unaryop_pack4->create("unaryop_pack4", opt, specializations, 1, 5);
     }
 
     return 0;

src/net.cpp

@@ -150,8 +150,17 @@ int Net::load_param(FILE* fp)
     blobs.resize((size_t)blob_count);
 
 #if NCNN_VULKAN
-    if (opt.use_vulkan_compute && !vkdev)
-        vkdev = get_default_gpu_device();
+    if (opt.use_vulkan_compute)
+    {
+        if (!vkdev) vkdev = get_default_gpu_device();
+
+        // sanitize use options
+        if (!vkdev->info.support_fp16_packed) opt.use_fp16_packed = false;
+        if (!vkdev->info.support_fp16_storage) opt.use_fp16_storage = false;
+        if (!vkdev->info.support_fp16_arithmetic) opt.use_fp16_arithmetic = false;
+        if (!vkdev->info.support_int8_storage) opt.use_int8_storage = false;
+        if (!vkdev->info.support_int8_arithmetic) opt.use_int8_arithmetic = false;
+    }
 #endif // NCNN_VULKAN
 
     ParamDict pd;
@@ -312,8 +321,17 @@ int Net::load_param_mem(const char* _mem)
     blobs.resize(blob_count);
 
 #if NCNN_VULKAN
-    if (opt.use_vulkan_compute && !vkdev)
-        vkdev = get_default_gpu_device();
+    if (opt.use_vulkan_compute)
+    {
+        if (!vkdev) vkdev = get_default_gpu_device();
+
+        // sanitize use options
+        if (!vkdev->info.support_fp16_packed) opt.use_fp16_packed = false;
+        if (!vkdev->info.support_fp16_storage) opt.use_fp16_storage = false;
+        if (!vkdev->info.support_fp16_arithmetic) opt.use_fp16_arithmetic = false;
+        if (!vkdev->info.support_int8_storage) opt.use_int8_storage = false;
+        if (!vkdev->info.support_int8_arithmetic) opt.use_int8_arithmetic = false;
+    }
 #endif // NCNN_VULKAN
 
     ParamDict pd;
@@ -478,8 +496,17 @@ int Net::load_param_bin(FILE* fp)
     blobs.resize(blob_count);
 
 #if NCNN_VULKAN
-    if (opt.use_vulkan_compute && !vkdev)
-        vkdev = get_default_gpu_device();
+    if (opt.use_vulkan_compute)
+    {
+        if (!vkdev) vkdev = get_default_gpu_device();
+
+        // sanitize use options
+        if (!vkdev->info.support_fp16_packed) opt.use_fp16_packed = false;
+        if (!vkdev->info.support_fp16_storage) opt.use_fp16_storage = false;
+        if (!vkdev->info.support_fp16_arithmetic) opt.use_fp16_arithmetic = false;
+        if (!vkdev->info.support_int8_storage) opt.use_int8_storage = false;
+        if (!vkdev->info.support_int8_arithmetic) opt.use_int8_arithmetic = false;
+    }
 #endif // NCNN_VULKAN
 
     ParamDict pd;
@@ -689,8 +716,17 @@ int Net::load_param(const unsigned char* _mem)
     blobs.resize(blob_count);
 
 #if NCNN_VULKAN
-    if (opt.use_vulkan_compute && !vkdev)
-        vkdev = get_default_gpu_device();
+    if (opt.use_vulkan_compute)
+    {
+        if (!vkdev) vkdev = get_default_gpu_device();
+
+        // sanitize use options
+        if (!vkdev->info.support_fp16_packed) opt.use_fp16_packed = false;
+        if (!vkdev->info.support_fp16_storage) opt.use_fp16_storage = false;
+        if (!vkdev->info.support_fp16_arithmetic) opt.use_fp16_arithmetic = false;
+        if (!vkdev->info.support_int8_storage) opt.use_int8_storage = false;
+        if (!vkdev->info.support_int8_arithmetic) opt.use_int8_arithmetic = false;
+    }
 #endif // NCNN_VULKAN
 
     ParamDict pd;
@@ -1028,7 +1064,7 @@ int Net::upload_model()
     {
         if (layers[i]->support_vulkan)
         {
-            int uret = layers[i]->upload_model(cmd);
+            int uret = layers[i]->upload_model(cmd, opt);
             if (uret != 0)
             {
                 fprintf(stderr, "layer upload_model %d failed\n", (int)i);
@@ -1044,7 +1080,7 @@ int Net::upload_model()
 
 int Net::create_pipeline()
 {
-    if (vkdev->info.support_fp16_packed || vkdev->info.support_fp16_storage)
+    if (opt.use_fp16_storage)
     {
         {
         cast_float32_to_float16 = ncnn::create_layer(ncnn::LayerType::Cast);
@@ -1375,7 +1411,7 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
 
                     // cast to fp16
                     VkMat bottom_blob_unpacked_fp16;
-                    if (vkdev->info.support_fp16_storage)
+                    if (opt.use_fp16_storage)
                     {
                         cast_float32_to_float16->forward(bottom_blob_unpacked, bottom_blob_unpacked_fp16, cmd, opt);
                     }
@@ -1477,7 +1513,7 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
 
                         // cast to fp16
                         VkMat bottom_blob_unpacked_fp16;
-                        if (vkdev->info.support_fp16_storage)
+                        if (opt.use_fp16_storage)
                         {
                             cast_float32_to_float16->forward(bottom_blob_unpacked, bottom_blob_unpacked_fp16, cmd, opt);
                         }
@@ -1606,7 +1642,7 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
 
                     // cast to fp32
                     VkMat bottom_blob_unpacked_fp32;
-                    if (vkdev->info.support_fp16_storage)
+                    if (opt.use_fp16_storage)
                     {
                         cast_float16_to_float32->forward(bottom_blob_unpacked, bottom_blob_unpacked_fp32, cmd, opt);
                     }
@@ -1741,7 +1777,7 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
                         packing_pack1->forward(bottom_blob, bottom_blob_unpacked, cmd, opt);
 
                         // cast to fp32
-                        if (vkdev->info.support_fp16_storage)
+                        if (opt.use_fp16_storage)
                         {
                             cast_float16_to_float32->forward(bottom_blob_unpacked, bottom_blobs_unpacked_fp32[i], cmd, opt);
                         }
@@ -1894,16 +1930,6 @@ void Extractor::set_num_threads(int num_threads)
     opt.num_threads = num_threads;
 }
 
-void Extractor::set_blob_allocator(Allocator* allocator)
-{
-    opt.blob_allocator = allocator;
-}
-
-void Extractor::set_workspace_allocator(Allocator* allocator)
-{
-    opt.workspace_allocator = allocator;
-}
-
 #if NCNN_VULKAN
 void Extractor::set_vulkan_compute(bool enable)
 {
@@ -1916,21 +1942,6 @@ void Extractor::set_vulkan_compute(bool enable)
         fprintf(stderr, "set_vulkan_compute failed, network use_vulkan_compute disabled\n");
     }
 }
-
-void Extractor::set_blob_vkallocator(VkAllocator* allocator)
-{
-    opt.blob_vkallocator = allocator;
-}
-
-void Extractor::set_workspace_vkallocator(VkAllocator* allocator)
-{
-    opt.workspace_vkallocator = allocator;
-}
-
-void Extractor::set_staging_vkallocator(VkAllocator* allocator)
-{
-    opt.staging_vkallocator = allocator;
-}
 #endif // NCNN_VULKAN
 
 #if NCNN_STRING
@@ -1993,7 +2004,7 @@ int Extractor::extract(int blob_index, Mat& feat)
 
                 // cast to fp32
                 VkMat feat_gpu_unpacked_fp32;
-                if (net->vkdev->info.support_fp16_packed || net->vkdev->info.support_fp16_storage)
+                if (opt.use_fp16_storage)
                 {
                     net->cast_float16_to_float32->forward(feat_gpu_unpacked, feat_gpu_unpacked_fp32, cmd, opt);
                 }

src/net.h

@@ -154,20 +154,8 @@ public:
     // default count is system depended
     void set_num_threads(int num_threads);
 
-    // set blob memory allocator
-    void set_blob_allocator(Allocator* allocator);
-
-    // set workspace memory allocator
-    void set_workspace_allocator(Allocator* allocator);
-
 #if NCNN_VULKAN
     void set_vulkan_compute(bool enable);
-
-    void set_blob_vkallocator(VkAllocator* allocator);
-
-    void set_workspace_vkallocator(VkAllocator* allocator);
-
-    void set_staging_vkallocator(VkAllocator* allocator);
 #endif // NCNN_VULKAN
 
 #if NCNN_STRING

src/pipeline.cpp

@@ -16,8 +16,9 @@
 #include <stdio.h>
 #include <math.h>
 #include <algorithm>
-#include "mat.h"
 #include <string>
+#include "mat.h"
+#include "option.h"
 
 namespace ncnn {
 
@@ -66,19 +67,19 @@ int Pipeline::create(VkShaderModule shader_module, const char* entry_name, const
     return 0;
 }
 
-int Pipeline::create(const char* _name, const std::vector<vk_specialization_type>& specializations, int binding_count, int push_constant_count)
+int Pipeline::create(const char* _name, const Option& opt, const std::vector<vk_specialization_type>& specializations, int binding_count, int push_constant_count)
 {
     std::string name = _name;
 
-    if (vkdev->info.support_fp16_arithmetic)
+    if (opt.use_fp16_arithmetic)
     {
         name += "_fp16a";
     }
-    else if (vkdev->info.support_fp16_storage)
+    else if (opt.use_fp16_storage)
     {
         name += "_fp16s";
     }
-    else if (vkdev->info.support_fp16_packed)
+    else if (opt.use_fp16_packed)
     {
         name += "_fp16p";
     }

src/pipeline.h

@@ -25,6 +25,7 @@
 namespace ncnn {
 
 #if NCNN_VULKAN
+class Option;
 class Pipeline
 {
 public:
@@ -39,7 +40,7 @@ public:
                const std::vector<vk_specialization_type>& specializations, int binding_count, int push_constant_count);
     int create(VkShaderModule shader_module, const char* entry_name,
                const std::vector<vk_specialization_type>& specializations, int binding_count, int push_constant_count);
-    int create(const char* name, const std::vector<vk_specialization_type>& specializations,
+    int create(const char* name, const Option& opt, const std::vector<vk_specialization_type>& specializations,
                int binding_count, int push_constant_count);
     void destroy();