enable pack8 shaders for layers without shape changes

src/layer/vulkan/absval_vulkan.cpp

@@ -24,6 +24,7 @@ AbsVal_vulkan::AbsVal_vulkan()
 
     pipeline_absval = 0;
     pipeline_absval_pack4 = 0;
+    pipeline_absval_pack8 = 0;
 }
 
 int AbsVal_vulkan::create_pipeline(const Option& opt)
@@ -44,6 +45,13 @@ int AbsVal_vulkan::create_pipeline(const Option& opt)
         pipeline_absval_pack4->create("absval_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_absval_pack8 = new Pipeline(vkdev);
+        pipeline_absval_pack8->set_optimal_local_size_xyz();
+        pipeline_absval_pack8->create("absval_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -55,6 +63,9 @@ int AbsVal_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_absval_pack4;
     pipeline_absval_pack4 = 0;
 
+    delete pipeline_absval_pack8;
+    pipeline_absval_pack8 = 0;
+
     return 0;
 }
 
@@ -72,7 +83,9 @@ int AbsVal_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_absval_pack4 : pipeline_absval;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_absval_pack8
+                             : elempack == 4 ? pipeline_absval_pack4
+                             : pipeline_absval;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/absval_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_absval;
     Pipeline* pipeline_absval_pack4;
+    Pipeline* pipeline_absval_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/cast_vulkan.cpp

@@ -24,8 +24,10 @@ Cast_vulkan::Cast_vulkan()
 
     pipeline_cast_fp32_to_fp16 = 0;
     pipeline_cast_fp32_to_fp16_pack4 = 0;
+    pipeline_cast_fp32_to_fp16_pack8 = 0;
     pipeline_cast_fp16_to_fp32 = 0;
     pipeline_cast_fp16_to_fp32_pack4 = 0;
+    pipeline_cast_fp16_to_fp32_pack8 = 0;
 }
 
 int Cast_vulkan::create_pipeline(const Option& opt)
@@ -47,6 +49,13 @@ int Cast_vulkan::create_pipeline(const Option& opt)
             pipeline_cast_fp32_to_fp16_pack4->set_optimal_local_size_xyz();
             pipeline_cast_fp32_to_fp16_pack4->create("cast_fp32_to_fp16_pack4", opt, specializations, 2, 10);
         }
+
+        // pack8
+        {
+            pipeline_cast_fp32_to_fp16_pack8 = new Pipeline(vkdev);
+            pipeline_cast_fp32_to_fp16_pack8->set_optimal_local_size_xyz();
+            pipeline_cast_fp32_to_fp16_pack8->create("cast_fp32_to_fp16_pack8", opt, specializations, 2, 10);
+        }
     }
 
     if (type_from == 2 && type_to == 1)
@@ -64,6 +73,13 @@ int Cast_vulkan::create_pipeline(const Option& opt)
             pipeline_cast_fp16_to_fp32_pack4->set_optimal_local_size_xyz();
             pipeline_cast_fp16_to_fp32_pack4->create("cast_fp16_to_fp32_pack4", opt, specializations, 2, 10);
         }
+
+        // pack8
+        {
+            pipeline_cast_fp16_to_fp32_pack8 = new Pipeline(vkdev);
+            pipeline_cast_fp16_to_fp32_pack8->set_optimal_local_size_xyz();
+            pipeline_cast_fp16_to_fp32_pack8->create("cast_fp16_to_fp32_pack8", opt, specializations, 2, 10);
+        }
     }
 
     return 0;
@@ -77,12 +93,18 @@ int Cast_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_cast_fp32_to_fp16_pack4;
     pipeline_cast_fp32_to_fp16_pack4 = 0;
 
+    delete pipeline_cast_fp32_to_fp16_pack8;
+    pipeline_cast_fp32_to_fp16_pack8 = 0;
+
     delete pipeline_cast_fp16_to_fp32;
     pipeline_cast_fp16_to_fp32 = 0;
 
     delete pipeline_cast_fp16_to_fp32_pack4;
     pipeline_cast_fp16_to_fp32_pack4 = 0;
 
+    delete pipeline_cast_fp16_to_fp32_pack8;
+    pipeline_cast_fp16_to_fp32_pack8 = 0;
+
     return 0;
 }
 
@@ -159,11 +181,15 @@ int Cast_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& c
 
     if (type_from == 1 && type_to == 2)
     {
-        pipeline = elempack == 4 ? pipeline_cast_fp32_to_fp16_pack4 : pipeline_cast_fp32_to_fp16;
+        pipeline = elempack == 8 ? pipeline_cast_fp32_to_fp16_pack8
+                 : elempack == 4 ? pipeline_cast_fp32_to_fp16_pack4
+                 : pipeline_cast_fp32_to_fp16;
     }
     if (type_from == 2 && type_to == 1)
     {
-        pipeline = elempack == 4 ? pipeline_cast_fp16_to_fp32_pack4 : pipeline_cast_fp16_to_fp32;
+        pipeline = elempack == 8 ? pipeline_cast_fp16_to_fp32_pack8
+                 : elempack == 4 ? pipeline_cast_fp16_to_fp32_pack4
+                 : pipeline_cast_fp16_to_fp32;
     }
 
     // TODO more cast type

src/layer/vulkan/cast_vulkan.h

@@ -33,8 +33,10 @@ public:
 public:
     Pipeline* pipeline_cast_fp32_to_fp16;
     Pipeline* pipeline_cast_fp32_to_fp16_pack4;
+    Pipeline* pipeline_cast_fp32_to_fp16_pack8;
     Pipeline* pipeline_cast_fp16_to_fp32;
     Pipeline* pipeline_cast_fp16_to_fp32_pack4;
+    Pipeline* pipeline_cast_fp16_to_fp32_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/clip_vulkan.cpp

@@ -24,6 +24,7 @@ Clip_vulkan::Clip_vulkan()
 
     pipeline_clip = 0;
     pipeline_clip_pack4 = 0;
+    pipeline_clip_pack8 = 0;
 }
 
 int Clip_vulkan::create_pipeline(const Option& opt)
@@ -46,6 +47,13 @@ int Clip_vulkan::create_pipeline(const Option& opt)
         pipeline_clip_pack4->create("clip_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_clip_pack8 = new Pipeline(vkdev);
+        pipeline_clip_pack8->set_optimal_local_size_xyz();
+        pipeline_clip_pack8->create("clip_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -57,6 +65,9 @@ int Clip_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_clip_pack4;
     pipeline_clip_pack4 = 0;
 
+    delete pipeline_clip_pack8;
+    pipeline_clip_pack8 = 0;
+
     return 0;
 }
 
@@ -74,7 +85,9 @@ int Clip_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const O
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_clip_pack4 : pipeline_clip;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_clip_pack8
+                             : elempack == 4 ? pipeline_clip_pack4
+                             : pipeline_clip;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/clip_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_clip;
     Pipeline* pipeline_clip_pack4;
+    Pipeline* pipeline_clip_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/dropout_vulkan.cpp

@@ -25,6 +25,7 @@ Dropout_vulkan::Dropout_vulkan()
 
     pipeline_dropout = 0;
     pipeline_dropout_pack4 = 0;
+    pipeline_dropout_pack8 = 0;
 }
 
 int Dropout_vulkan::create_pipeline(const Option& opt)
@@ -46,6 +47,13 @@ int Dropout_vulkan::create_pipeline(const Option& opt)
         pipeline_dropout_pack4->create("dropout_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_dropout_pack8 = new Pipeline(vkdev);
+        pipeline_dropout_pack8->set_optimal_local_size_xyz();
+        pipeline_dropout_pack8->create("dropout_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -57,6 +65,9 @@ int Dropout_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_dropout_pack4;
     pipeline_dropout_pack4 = 0;
 
+    delete pipeline_dropout_pack8;
+    pipeline_dropout_pack8 = 0;
+
     return 0;
 }
 
@@ -79,7 +90,9 @@ int Dropout_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, cons
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_dropout_pack4 : pipeline_dropout;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_dropout_pack8
+                             : elempack == 4 ? pipeline_dropout_pack4
+                             : pipeline_dropout;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/dropout_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_dropout;
     Pipeline* pipeline_dropout_pack4;
+    Pipeline* pipeline_dropout_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/eltwise_vulkan.cpp

@@ -27,6 +27,8 @@ Eltwise_vulkan::Eltwise_vulkan()
     pipeline_eltwise[1] = 0;
     pipeline_eltwise_pack4[0] = 0;
     pipeline_eltwise_pack4[1] = 0;
+    pipeline_eltwise_pack8[0] = 0;
+    pipeline_eltwise_pack8[1] = 0;
 }
 
 int Eltwise_vulkan::create_pipeline(const Option& opt)
@@ -55,6 +57,16 @@ int Eltwise_vulkan::create_pipeline(const Option& opt)
         pipeline_eltwise_pack4[1]->create("eltwise_pack4", opt, specializations, 3, 5+2);
     }
 
+    // pack8
+    {
+        pipeline_eltwise_pack8[0] = new Pipeline(vkdev);
+        pipeline_eltwise_pack8[0]->set_optimal_local_size_xyz();
+        pipeline_eltwise_pack8[0]->create("eltwise_pack8", opt, specializations, 3, 5+2);
+        pipeline_eltwise_pack8[1] = new Pipeline(vkdev);
+        pipeline_eltwise_pack8[1]->set_optimal_local_size_xyz();
+        pipeline_eltwise_pack8[1]->create("eltwise_pack8", opt, specializations, 3, 5+2);
+    }
+
     return 0;
 }
 
@@ -70,6 +82,11 @@ int Eltwise_vulkan::destroy_pipeline(const Option& /*opt*/)
     pipeline_eltwise_pack4[0] = 0;
     pipeline_eltwise_pack4[1] = 0;
 
+    delete pipeline_eltwise_pack8[0];
+    delete pipeline_eltwise_pack8[1];
+    pipeline_eltwise_pack8[0] = 0;
+    pipeline_eltwise_pack8[1] = 0;
+
     return 0;
 }
 
@@ -103,7 +120,9 @@ int Eltwise_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector<
     constants[5].f = coeffs.w == 0 ? 1.f : coeffs[0];
     constants[6].f = coeffs.w == 0 ? 1.f : coeffs[1];
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_eltwise_pack4[1] : pipeline_eltwise[1];
+    const Pipeline* pipeline = elempack == 8 ? pipeline_eltwise_pack8[1]
+                             : elempack == 4 ? pipeline_eltwise_pack4[1]
+                             : pipeline_eltwise[1];
 
     cmd.record_pipeline(pipeline, bindings, constants, top_blob);
 
@@ -123,7 +142,9 @@ int Eltwise_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector<
         constants[5].f = 1.f;
         constants[6].f = coeffs.w == 0 ? 1 : coeffs[b];
 
-        const Pipeline* pipeline = elempack == 4 ? pipeline_eltwise_pack4[b%2] : pipeline_eltwise[b%2];
+        const Pipeline* pipeline = elempack == 8 ? pipeline_eltwise_pack8[b%2]
+                                 : elempack == 4 ? pipeline_eltwise_pack4[b%2]
+                                 : pipeline_eltwise[b%2];
 
         cmd.record_pipeline(pipeline, bindings, constants, top_blob);
     }

src/layer/vulkan/eltwise_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_eltwise[2];
     Pipeline* pipeline_eltwise_pack4[2];
+    Pipeline* pipeline_eltwise_pack8[2];
 };
 
 } // namespace ncnn

src/layer/vulkan/hardsigmoid_vulkan.cpp

@@ -24,6 +24,7 @@ HardSigmoid_vulkan::HardSigmoid_vulkan()
 
     pipeline_hardsigmoid = 0;
     pipeline_hardsigmoid_pack4 = 0;
+    pipeline_hardsigmoid_pack8 = 0;
 }
 
 int HardSigmoid_vulkan::create_pipeline(const Option& opt)
@@ -46,6 +47,13 @@ int HardSigmoid_vulkan::create_pipeline(const Option& opt)
         pipeline_hardsigmoid_pack4->create("hardsigmoid_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_hardsigmoid_pack8 = new Pipeline(vkdev);
+        pipeline_hardsigmoid_pack8->set_optimal_local_size_xyz();
+        pipeline_hardsigmoid_pack8->create("hardsigmoid_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -57,6 +65,9 @@ int HardSigmoid_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_hardsigmoid_pack4;
     pipeline_hardsigmoid_pack4 = 0;
 
+    delete pipeline_hardsigmoid_pack8;
+    pipeline_hardsigmoid_pack8 = 0;
+
     return 0;
 }
 
@@ -74,7 +85,9 @@ int HardSigmoid_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd,
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_hardsigmoid_pack4 : pipeline_hardsigmoid;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_hardsigmoid_pack8
+                             : elempack == 4 ? pipeline_hardsigmoid_pack4
+                             : pipeline_hardsigmoid;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/hardsigmoid_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_hardsigmoid;
     Pipeline* pipeline_hardsigmoid_pack4;
+    Pipeline* pipeline_hardsigmoid_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/hardswish_vulkan.cpp

@@ -24,6 +24,7 @@ HardSwish_vulkan::HardSwish_vulkan()
 
     pipeline_hardswish = 0;
     pipeline_hardswish_pack4 = 0;
+    pipeline_hardswish_pack8 = 0;
 }
 
 int HardSwish_vulkan::create_pipeline(const Option& opt)
@@ -46,6 +47,13 @@ int HardSwish_vulkan::create_pipeline(const Option& opt)
         pipeline_hardswish_pack4->create("hardswish_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_hardswish_pack8 = new Pipeline(vkdev);
+        pipeline_hardswish_pack8->set_optimal_local_size_xyz();
+        pipeline_hardswish_pack8->create("hardswish_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -57,6 +65,9 @@ int HardSwish_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_hardswish_pack4;
     pipeline_hardswish_pack4 = 0;
 
+    delete pipeline_hardswish_pack8;
+    pipeline_hardswish_pack8 = 0;
+
     return 0;
 }
 
@@ -74,7 +85,9 @@ int HardSwish_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, co
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_hardswish_pack4 : pipeline_hardswish;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_hardswish_pack8
+                             : elempack == 4 ? pipeline_hardswish_pack4
+                             : pipeline_hardswish;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/hardswish_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_hardswish;
     Pipeline* pipeline_hardswish_pack4;
+    Pipeline* pipeline_hardswish_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/lrn_vulkan.cpp

@@ -29,6 +29,10 @@ LRN_vulkan::LRN_vulkan()
     pipeline_lrn_norm_across_channel_pack4 = 0;
     pipeline_lrn_square_pad_within_channel_pack4 = 0;
     pipeline_lrn_norm_within_channel_pack4 = 0;
+    pipeline_lrn_square_pad_across_channel_pack8 = 0;
+    pipeline_lrn_norm_across_channel_pack8 = 0;
+    pipeline_lrn_square_pad_within_channel_pack8 = 0;
+    pipeline_lrn_norm_within_channel_pack8 = 0;
 }
 
 int LRN_vulkan::create_pipeline(const Option& opt)
@@ -67,6 +71,20 @@ int LRN_vulkan::create_pipeline(const Option& opt)
             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", opt, specializations, 2, 10);
         }
+
+        // pack8
+        if (region_type == 0)
+        {
+            pipeline_lrn_square_pad_across_channel_pack8 = new Pipeline(vkdev);
+            pipeline_lrn_square_pad_across_channel_pack8->set_optimal_local_size_xyz();
+            pipeline_lrn_square_pad_across_channel_pack8->create("lrn_square_pad_across_channel_pack8", opt, specializations, 2, 10);
+        }
+        if (region_type == 1)
+        {
+            pipeline_lrn_square_pad_within_channel_pack8 = new Pipeline(vkdev);
+            pipeline_lrn_square_pad_within_channel_pack8->set_optimal_local_size_xyz();
+            pipeline_lrn_square_pad_within_channel_pack8->create("lrn_square_pad_within_channel_pack8", opt, specializations, 2, 10);
+        }
     }
 
     {
@@ -95,6 +113,20 @@ int LRN_vulkan::create_pipeline(const Option& opt)
             pipeline_lrn_norm_within_channel_pack4->set_optimal_local_size_xyz();
             pipeline_lrn_norm_within_channel_pack4->create("lrn_norm_within_channel_pack4", opt, specializations, 2, 10);
         }
+
+        // pack8
+        if (region_type == 0)
+        {
+            pipeline_lrn_norm_across_channel_pack8 = new Pipeline(vkdev);
+            pipeline_lrn_norm_across_channel_pack8->set_optimal_local_size_xyz();
+            pipeline_lrn_norm_across_channel_pack8->create("lrn_norm_across_channel_pack8", opt, specializations, 2, 10);
+        }
+        if (region_type == 1)
+        {
+            pipeline_lrn_norm_within_channel_pack8 = new Pipeline(vkdev);
+            pipeline_lrn_norm_within_channel_pack8->set_optimal_local_size_xyz();
+            pipeline_lrn_norm_within_channel_pack8->create("lrn_norm_within_channel_pack8", opt, specializations, 2, 10);
+        }
     }
 
     return 0;
@@ -120,6 +152,18 @@ int LRN_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_lrn_norm_within_channel_pack4;
     pipeline_lrn_norm_within_channel_pack4 = 0;
 
+    delete pipeline_lrn_square_pad_across_channel_pack8;
+    pipeline_lrn_square_pad_across_channel_pack8 = 0;
+
+    delete pipeline_lrn_norm_across_channel_pack8;
+    pipeline_lrn_norm_across_channel_pack8 = 0;
+
+    delete pipeline_lrn_square_pad_within_channel_pack8;
+    pipeline_lrn_square_pad_within_channel_pack8 = 0;
+
+    delete pipeline_lrn_norm_within_channel_pack8;
+    pipeline_lrn_norm_within_channel_pack8 = 0;
+
     return 0;
 }
 
@@ -167,7 +211,12 @@ int LRN_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Op
     constants[9].i = square_workspace.cstep;
 
     const Pipeline* pipeline = 0;
-    if (elempack == 4)
+    if (elempack == 8)
+    {
+        if (region_type == 0) pipeline = pipeline_lrn_square_pad_across_channel_pack8;
+        if (region_type == 1) pipeline = pipeline_lrn_square_pad_within_channel_pack8;
+    }
+    else if (elempack == 4)
     {
         if (region_type == 0) pipeline = pipeline_lrn_square_pad_across_channel_pack4;
         if (region_type == 1) pipeline = pipeline_lrn_square_pad_within_channel_pack4;
@@ -199,7 +248,12 @@ int LRN_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Op
     constants[9].i = bottom_top_blob.cstep;
 
     const Pipeline* pipeline = 0;
-    if (elempack == 4)
+    if (elempack == 8)
+    {
+        if (region_type == 0) pipeline = pipeline_lrn_norm_across_channel_pack8;
+        if (region_type == 1) pipeline = pipeline_lrn_norm_within_channel_pack8;
+    }
+    else if (elempack == 4)
     {
         if (region_type == 0) pipeline = pipeline_lrn_norm_across_channel_pack4;
         if (region_type == 1) pipeline = pipeline_lrn_norm_within_channel_pack4;

src/layer/vulkan/lrn_vulkan.h

@@ -37,6 +37,10 @@ public:
     Pipeline* pipeline_lrn_norm_across_channel_pack4;
     Pipeline* pipeline_lrn_square_pad_within_channel_pack4;
     Pipeline* pipeline_lrn_norm_within_channel_pack4;
+    Pipeline* pipeline_lrn_square_pad_across_channel_pack8;
+    Pipeline* pipeline_lrn_norm_across_channel_pack8;
+    Pipeline* pipeline_lrn_square_pad_within_channel_pack8;
+    Pipeline* pipeline_lrn_norm_within_channel_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/normalize_vulkan.cpp

@@ -33,6 +33,12 @@ Normalize_vulkan::Normalize_vulkan()
     pipeline_normalize_reduce_sum4_fp32_pack4[1] = 0;
     pipeline_normalize_coeffs_pack4 = 0;
     pipeline_normalize_norm_pack4 = 0;
+
+    pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8 = 0;
+    pipeline_normalize_reduce_sum4_fp32_pack8[0] = 0;
+    pipeline_normalize_reduce_sum4_fp32_pack8[1] = 0;
+    pipeline_normalize_coeffs_pack8 = 0;
+    pipeline_normalize_norm_pack8 = 0;
 }
 
 int Normalize_vulkan::create_pipeline(const Option& opt)
@@ -65,6 +71,18 @@ int Normalize_vulkan::create_pipeline(const Option& opt)
         pipeline_normalize_reduce_sum4_fp32_pack4[1] = new Pipeline(vkdev);
         pipeline_normalize_reduce_sum4_fp32_pack4[1]->set_optimal_local_size_xyz();
         pipeline_normalize_reduce_sum4_fp32_pack4[1]->create("normalize_reduce_sum4_fp32_pack4", opt, specializations, 2, 6);
+
+        // pack8
+        pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8 = new Pipeline(vkdev);
+        pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8->set_optimal_local_size_xyz();
+        pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8->create("normalize_reduce_sum4_fp16_to_fp32_pack8", opt, specializations, 2, 6);
+
+        pipeline_normalize_reduce_sum4_fp32_pack8[0] = new Pipeline(vkdev);
+        pipeline_normalize_reduce_sum4_fp32_pack8[0]->set_optimal_local_size_xyz();
+        pipeline_normalize_reduce_sum4_fp32_pack8[0]->create("normalize_reduce_sum4_fp32_pack8", opt, specializations, 2, 6);
+        pipeline_normalize_reduce_sum4_fp32_pack8[1] = new Pipeline(vkdev);
+        pipeline_normalize_reduce_sum4_fp32_pack8[1]->set_optimal_local_size_xyz();
+        pipeline_normalize_reduce_sum4_fp32_pack8[1]->create("normalize_reduce_sum4_fp32_pack8", opt, specializations, 2, 6);
     }
 
     {
@@ -79,6 +97,10 @@ int Normalize_vulkan::create_pipeline(const Option& opt)
         pipeline_normalize_coeffs_pack4 = new Pipeline(vkdev);
         pipeline_normalize_coeffs_pack4->set_optimal_local_size_xyz();
         pipeline_normalize_coeffs_pack4->create("normalize_coeffs_pack4", opt, specializations, 2, 3);
+
+        pipeline_normalize_coeffs_pack8 = new Pipeline(vkdev);
+        pipeline_normalize_coeffs_pack8->set_optimal_local_size_xyz();
+        pipeline_normalize_coeffs_pack8->create("normalize_coeffs_pack8", opt, specializations, 2, 3);
     }
 
     {
@@ -95,6 +117,10 @@ int Normalize_vulkan::create_pipeline(const Option& opt)
         pipeline_normalize_norm_pack4 = new Pipeline(vkdev);
         pipeline_normalize_norm_pack4->set_optimal_local_size_xyz();
         pipeline_normalize_norm_pack4->create("normalize_norm_pack4", opt, specializations, 3, 5);
+
+        pipeline_normalize_norm_pack8 = new Pipeline(vkdev);
+        pipeline_normalize_norm_pack8->set_optimal_local_size_xyz();
+        pipeline_normalize_norm_pack8->create("normalize_norm_pack8", opt, specializations, 3, 5);
     }
 
     return 0;
@@ -118,18 +144,32 @@ int Normalize_vulkan::destroy_pipeline(const Option& /*opt*/)
     pipeline_normalize_reduce_sum4_fp32_pack4[0] = 0;
     pipeline_normalize_reduce_sum4_fp32_pack4[1] = 0;
 
+    delete pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8;
+    pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8 = 0;
+
+    delete pipeline_normalize_reduce_sum4_fp32_pack8[0];
+    delete pipeline_normalize_reduce_sum4_fp32_pack8[1];
+    pipeline_normalize_reduce_sum4_fp32_pack8[0] = 0;
+    pipeline_normalize_reduce_sum4_fp32_pack8[1] = 0;
+
     delete pipeline_normalize_coeffs;
     pipeline_normalize_coeffs = 0;
 
     delete pipeline_normalize_coeffs_pack4;
     pipeline_normalize_coeffs_pack4 = 0;
 
+    delete pipeline_normalize_coeffs_pack8;
+    pipeline_normalize_coeffs_pack8 = 0;
+
     delete pipeline_normalize_norm;
     pipeline_normalize_norm = 0;
 
     delete pipeline_normalize_norm_pack4;
     pipeline_normalize_norm_pack4 = 0;
 
+    delete pipeline_normalize_norm_pack8;
+    pipeline_normalize_norm_pack8 = 0;
+
     return 0;
 }
 
@@ -141,16 +181,6 @@ int Normalize_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
         Mat scale_data4(4);
         scale_data4.fill(scale_data[0]);
         cmd.record_upload(scale_data4, scale_data_gpu, opt);
-
-        Mat scale_data_pack4;
-        convert_packing(scale_data4, scale_data_pack4, 4);
-        cmd.record_upload(scale_data_pack4, scale_data_gpu_pack4, opt);
-    }
-    else if (scale_data_size % 4 == 0)
-    {
-        Mat scale_data_pack4;
-        convert_packing(scale_data, scale_data_pack4, 4);
-        cmd.record_upload(scale_data_pack4, scale_data_gpu_pack4, opt);
     }
     else
     {
@@ -209,7 +239,9 @@ int Normalize_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, co
         constants[4].i = sqsum_workspace.c;
         constants[5].i = sqsum_workspace.cstep;
 
-        const Pipeline* pipeline = elempack == 4 ? pipeline_normalize_reduce_sum4_fp16_to_fp32_pack4 : pipeline_normalize_reduce_sum4_fp16_to_fp32;
+        const Pipeline* pipeline = elempack == 8 ? pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8
+                                 : elempack == 4 ? pipeline_normalize_reduce_sum4_fp16_to_fp32_pack4
+                                 : pipeline_normalize_reduce_sum4_fp16_to_fp32;
 
         cmd.record_pipeline(pipeline, bindings, constants, sqsum_workspace);
         }
@@ -257,7 +289,9 @@ int Normalize_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, co
         constants[4].i = sqsum_workspace_reduced.c;
         constants[5].i = sqsum_workspace_reduced.cstep;
 
-        const Pipeline* pipeline = elempack == 4 ? pipeline_normalize_reduce_sum4_fp32_pack4[pb%2] : pipeline_normalize_reduce_sum4_fp32[pb%2];
+        const Pipeline* pipeline = elempack == 8 ? pipeline_normalize_reduce_sum4_fp32_pack8[pb%2]
+                                 : elempack == 4 ? pipeline_normalize_reduce_sum4_fp32_pack4[pb%2]
+                                 : pipeline_normalize_reduce_sum4_fp32[pb%2];
 
         cmd.record_pipeline(pipeline, bindings, constants, sqsum_workspace_reduced);
 
@@ -281,7 +315,9 @@ int Normalize_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, co
         constants[1].i = sqsum_workspace.c;
         constants[2].i = sqsum_workspace.cstep;
 
-        const Pipeline* pipeline = elempack == 4 ? pipeline_normalize_coeffs_pack4 : pipeline_normalize_coeffs;
+        const Pipeline* pipeline = elempack == 8 ? pipeline_normalize_coeffs_pack8
+                                 : elempack == 4 ? pipeline_normalize_coeffs_pack4
+                                 : pipeline_normalize_coeffs;
 
         cmd.record_pipeline(pipeline, bindings, constants, coeffs_workspace);
     }
@@ -291,7 +327,7 @@ int Normalize_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, co
         std::vector<VkMat> bindings(3);
         bindings[0] = bottom_top_blob;
         bindings[1] = coeffs_workspace;
-        bindings[2] = (scale_data_size == 1 && scale_data[0] == 1.f) ? coeffs_workspace : elempack == 4 ? scale_data_gpu_pack4 : scale_data_gpu;
+        bindings[2] = (scale_data_size == 1 && scale_data[0] == 1.f) ? coeffs_workspace : scale_data_gpu;
 
         std::vector<vk_constant_type> constants(5);
         constants[0].i = bottom_top_blob.dims;
@@ -300,7 +336,9 @@ int Normalize_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, co
         constants[3].i = bottom_top_blob.c;
         constants[4].i = bottom_top_blob.cstep;
 
-        const Pipeline* pipeline = elempack == 4 ? pipeline_normalize_norm_pack4 : pipeline_normalize_norm;
+        const Pipeline* pipeline = elempack == 8 ? pipeline_normalize_norm_pack8
+                                 : elempack == 4 ? pipeline_normalize_norm_pack4
+                                 : pipeline_normalize_norm;
 
         cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
     }

src/layer/vulkan/normalize_vulkan.h

@@ -39,11 +39,15 @@ public:
     Pipeline* pipeline_normalize_coeffs;
     Pipeline* pipeline_normalize_norm;
 
-    VkMat scale_data_gpu_pack4;
     Pipeline* pipeline_normalize_reduce_sum4_fp16_to_fp32_pack4;
     Pipeline* pipeline_normalize_reduce_sum4_fp32_pack4[2];
     Pipeline* pipeline_normalize_coeffs_pack4;
     Pipeline* pipeline_normalize_norm_pack4;
+
+    Pipeline* pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8;
+    Pipeline* pipeline_normalize_reduce_sum4_fp32_pack8[2];
+    Pipeline* pipeline_normalize_coeffs_pack8;
+    Pipeline* pipeline_normalize_norm_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/padding_vulkan.cpp

@@ -24,6 +24,7 @@ Padding_vulkan::Padding_vulkan()
 
     pipeline_padding = 0;
     pipeline_padding_pack4 = 0;
+    pipeline_padding_pack8 = 0;
 }
 
 int Padding_vulkan::create_pipeline(const Option& opt)
@@ -47,6 +48,13 @@ int Padding_vulkan::create_pipeline(const Option& opt)
         pipeline_padding_pack4->create("padding_pack4", opt, specializations, 3, 12);
     }
 
+    // pack8
+    {
+        pipeline_padding_pack8 = new Pipeline(vkdev);
+        pipeline_padding_pack8->set_optimal_local_size_xyz();
+        pipeline_padding_pack8->create("padding_pack8", opt, specializations, 3, 12);
+    }
+
     return 0;
 }
 
@@ -58,6 +66,9 @@ int Padding_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_padding_pack4;
     pipeline_padding_pack4 = 0;
 
+    delete pipeline_padding_pack8;
+    pipeline_padding_pack8 = 0;
+
     return 0;
 }
 
@@ -66,19 +77,7 @@ int Padding_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
     if (per_channel_pad_data_size == 0)
         return 0;
 
-    // pack1
-    if (per_channel_pad_data_size % 4 != 0)
-    {
-        cmd.record_upload(per_channel_pad_data, per_channel_pad_data_gpu, opt);
-    }
-
-    // pack4
-    if (per_channel_pad_data_size % 4 == 0)
-    {
-        Mat per_channel_pad_data_pack4;
-        convert_packing(per_channel_pad_data, per_channel_pad_data_pack4, 4);
-        cmd.record_upload(per_channel_pad_data_pack4, per_channel_pad_data_gpu_pack4, opt);
-    }
+    cmd.record_upload(per_channel_pad_data, per_channel_pad_data_gpu, opt);
 
     return 0;
 }
@@ -109,7 +108,7 @@ int Padding_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
     std::vector<VkMat> bindings(3);
     bindings[0] = bottom_blob;
     bindings[1] = top_blob;
-    bindings[2] = per_channel_pad_data_size ? (elempack == 4 ? per_channel_pad_data_gpu_pack4 : per_channel_pad_data_gpu) : top_blob;// TODO use dummy buffer
+    bindings[2] = per_channel_pad_data_size ? per_channel_pad_data_gpu : top_blob;// TODO use dummy buffer
 
     std::vector<vk_constant_type> constants(12);
     constants[0].i = bottom_blob.dims;
@@ -125,7 +124,9 @@ int Padding_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
     constants[10].i = left;
     constants[11].i = top;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_padding_pack4 : pipeline_padding;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_padding_pack8
+                             : elempack == 4 ? pipeline_padding_pack4
+                             : pipeline_padding;
 
     cmd.record_pipeline(pipeline, bindings, constants, top_blob);
 
@@ -176,7 +177,7 @@ int Padding_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector<
     std::vector<VkMat> bindings(3);
     bindings[0] = bottom_blob;
     bindings[1] = top_blob;
-    bindings[2] = per_channel_pad_data_size ? (elempack == 4 ? per_channel_pad_data_gpu_pack4 : per_channel_pad_data_gpu) : top_blob;// TODO use dummy buffer
+    bindings[2] = per_channel_pad_data_size ? per_channel_pad_data_gpu : top_blob;// TODO use dummy buffer
 
     std::vector<vk_constant_type> constants(12);
     constants[0].i = bottom_blob.dims;
@@ -192,7 +193,9 @@ int Padding_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector<
     constants[10].i = _left;
     constants[11].i = _top;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_padding_pack4 : pipeline_padding;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_padding_pack8
+                             : elempack == 4 ? pipeline_padding_pack4
+                             : pipeline_padding;
 
     cmd.record_pipeline(pipeline, bindings, constants, top_blob);
 

src/layer/vulkan/padding_vulkan.h

@@ -37,9 +37,8 @@ public:
 public:
     VkMat per_channel_pad_data_gpu;
     Pipeline* pipeline_padding;
-
-    VkMat per_channel_pad_data_gpu_pack4;
     Pipeline* pipeline_padding_pack4;
+    Pipeline* pipeline_padding_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/pooling_vulkan.cpp

@@ -27,9 +27,11 @@ Pooling_vulkan::Pooling_vulkan()
 
     padding = 0;
     pipeline_pooling = 0;
-    pipeline_pooling_global = 0;
     pipeline_pooling_pack4 = 0;
+    pipeline_pooling_pack8 = 0;
+    pipeline_pooling_global = 0;
     pipeline_pooling_global_pack4 = 0;
+    pipeline_pooling_global_pack8 = 0;
 }
 
 int Pooling_vulkan::create_pipeline(const Option& opt)
@@ -59,34 +61,6 @@ int Pooling_vulkan::create_pipeline(const Option& opt)
         padding->create_pipeline(opt);
     }
 
-    std::vector<vk_specialization_type> specializations(12);
-    specializations[0].i = pooling_type;
-    specializations[1].i = kernel_w;
-    specializations[2].i = kernel_h;
-    specializations[3].i = stride_w;
-    specializations[4].i = stride_h;
-    specializations[5].i = pad_left;
-    specializations[6].i = pad_right;
-    specializations[7].i = pad_top;
-    specializations[8].i = pad_bottom;
-    specializations[9].i = global_pooling;
-    specializations[10].i = pad_mode;
-    specializations[11].i = avgpool_count_include_pad;
-
-    // pack1
-    {
-        pipeline_pooling = new Pipeline(vkdev);
-        pipeline_pooling->set_optimal_local_size_xyz();
-        pipeline_pooling->create("pooling", opt, specializations, 2, 12);
-    }
-
-    // pack4
-    {
-        pipeline_pooling_pack4 = new Pipeline(vkdev);
-        pipeline_pooling_pack4->set_optimal_local_size_xyz();
-        pipeline_pooling_pack4->create("pooling_pack4", opt, specializations, 2, 12);
-    }
-
     if (global_pooling)
     {
         std::vector<vk_specialization_type> specializations(1);
@@ -105,6 +79,50 @@ int Pooling_vulkan::create_pipeline(const Option& opt)
             pipeline_pooling_global_pack4->set_optimal_local_size_xyz(256, 1, 1);
             pipeline_pooling_global_pack4->create("pooling_global_pack4", opt, specializations, 2, 12);
         }
+
+        // pack8
+        {
+            pipeline_pooling_global_pack8 = new Pipeline(vkdev);
+            pipeline_pooling_global_pack8->set_optimal_local_size_xyz(256, 1, 1);
+            pipeline_pooling_global_pack8->create("pooling_global_pack8", opt, specializations, 2, 12);
+        }
+    }
+    else
+    {
+        std::vector<vk_specialization_type> specializations(12);
+        specializations[0].i = pooling_type;
+        specializations[1].i = kernel_w;
+        specializations[2].i = kernel_h;
+        specializations[3].i = stride_w;
+        specializations[4].i = stride_h;
+        specializations[5].i = pad_left;
+        specializations[6].i = pad_right;
+        specializations[7].i = pad_top;
+        specializations[8].i = pad_bottom;
+        specializations[9].i = global_pooling;
+        specializations[10].i = pad_mode;
+        specializations[11].i = avgpool_count_include_pad;
+
+        // pack1
+        {
+            pipeline_pooling = new Pipeline(vkdev);
+            pipeline_pooling->set_optimal_local_size_xyz();
+            pipeline_pooling->create("pooling", opt, specializations, 2, 12);
+        }
+
+        // pack4
+        {
+            pipeline_pooling_pack4 = new Pipeline(vkdev);
+            pipeline_pooling_pack4->set_optimal_local_size_xyz();
+            pipeline_pooling_pack4->create("pooling_pack4", opt, specializations, 2, 12);
+        }
+
+        // pack8
+        {
+            pipeline_pooling_pack8 = new Pipeline(vkdev);
+            pipeline_pooling_pack8->set_optimal_local_size_xyz();
+            pipeline_pooling_pack8->create("pooling_pack8", opt, specializations, 2, 12);
+        }
     }
 
     return 0;
@@ -125,12 +143,18 @@ int Pooling_vulkan::destroy_pipeline(const Option& opt)
     delete pipeline_pooling_pack4;
     pipeline_pooling_pack4 = 0;
 
+    delete pipeline_pooling_pack8;
+    pipeline_pooling_pack8 = 0;
+
     delete pipeline_pooling_global;
     pipeline_pooling_global = 0;
 
     delete pipeline_pooling_global_pack4;
     pipeline_pooling_global_pack4 = 0;
 
+    delete pipeline_pooling_global_pack8;
+    pipeline_pooling_global_pack8 = 0;
+
     return 0;
 }
 
@@ -166,7 +190,9 @@ int Pooling_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
         constants[10].i = 0;
         constants[11].i = 0;
 
-        const Pipeline* pipeline = elempack == 4 ? pipeline_pooling_global_pack4 : pipeline_pooling_global;
+        const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_pack8
+                                 : elempack == 4 ? pipeline_pooling_global_pack4
+                                 : pipeline_pooling_global;
 
         cmd.record_pipeline(pipeline, bindings, constants, top_blob);
 
@@ -298,7 +324,9 @@ int Pooling_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
     constants[10].i = wtailpad;
     constants[11].i = htailpad;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_pooling_pack4 : pipeline_pooling;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_pack8
+                             : elempack == 4 ? pipeline_pooling_pack4
+                             : pipeline_pooling;
 
     cmd.record_pipeline(pipeline, bindings, constants, top_blob);
 

src/layer/vulkan/pooling_vulkan.h

@@ -34,9 +34,11 @@ public:
     ncnn::Layer* padding;
 
     Pipeline* pipeline_pooling;
-    Pipeline* pipeline_pooling_global;
     Pipeline* pipeline_pooling_pack4;
+    Pipeline* pipeline_pooling_pack8;
+    Pipeline* pipeline_pooling_global;
     Pipeline* pipeline_pooling_global_pack4;
+    Pipeline* pipeline_pooling_global_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/relu_vulkan.cpp

@@ -24,6 +24,7 @@ ReLU_vulkan::ReLU_vulkan()
 
     pipeline_relu = 0;
     pipeline_relu_pack4 = 0;
+    pipeline_relu_pack8 = 0;
 }
 
 int ReLU_vulkan::create_pipeline(const Option& opt)
@@ -45,6 +46,13 @@ int ReLU_vulkan::create_pipeline(const Option& opt)
         pipeline_relu_pack4->create("relu_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_relu_pack8 = new Pipeline(vkdev);
+        pipeline_relu_pack8->set_optimal_local_size_xyz();
+        pipeline_relu_pack8->create("relu_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -56,6 +64,9 @@ int ReLU_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_relu_pack4;
     pipeline_relu_pack4 = 0;
 
+    delete pipeline_relu_pack8;
+    pipeline_relu_pack8 = 0;
+
     return 0;
 }
 
@@ -73,7 +84,9 @@ int ReLU_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const O
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_relu_pack4 : pipeline_relu;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_relu_pack8
+                             : elempack == 4 ? pipeline_relu_pack4
+                             : pipeline_relu;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/relu_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_relu;
     Pipeline* pipeline_relu_pack4;
+    Pipeline* pipeline_relu_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/sigmoid_vulkan.cpp

@@ -25,6 +25,7 @@ Sigmoid_vulkan::Sigmoid_vulkan()
 
     pipeline_sigmoid = 0;
     pipeline_sigmoid_pack4 = 0;
+    pipeline_sigmoid_pack8 = 0;
 }
 
 int Sigmoid_vulkan::create_pipeline(const Option& opt)
@@ -45,6 +46,13 @@ int Sigmoid_vulkan::create_pipeline(const Option& opt)
         pipeline_sigmoid_pack4->create("sigmoid_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_sigmoid_pack8 = new Pipeline(vkdev);
+        pipeline_sigmoid_pack8->set_optimal_local_size_xyz();
+        pipeline_sigmoid_pack8->create("sigmoid_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -56,6 +64,9 @@ int Sigmoid_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_sigmoid_pack4;
     pipeline_sigmoid_pack4 = 0;
 
+    delete pipeline_sigmoid_pack8;
+    pipeline_sigmoid_pack8 = 0;
+
     return 0;
 }
 
@@ -73,7 +84,9 @@ int Sigmoid_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, cons
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_sigmoid_pack4 : pipeline_sigmoid;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_sigmoid_pack8
+                             : elempack == 4 ? pipeline_sigmoid_pack4
+                             : pipeline_sigmoid;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/sigmoid_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_sigmoid;
     Pipeline* pipeline_sigmoid_pack4;
+    Pipeline* pipeline_sigmoid_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/softmax_vulkan.cpp

@@ -34,6 +34,11 @@ Softmax_vulkan::Softmax_vulkan()
     pipeline_softmax_exp_sub_max_pack4 = 0;
     pipeline_softmax_reduce_sum_pack4 = 0;
     pipeline_softmax_div_sum_pack4 = 0;
+
+    pipeline_softmax_reduce_max_pack8 = 0;
+    pipeline_softmax_exp_sub_max_pack8 = 0;
+    pipeline_softmax_reduce_sum_pack8 = 0;
+    pipeline_softmax_div_sum_pack8 = 0;
 }
 
 int Softmax_vulkan::create_pipeline(const Option& opt)
@@ -77,6 +82,24 @@ int Softmax_vulkan::create_pipeline(const Option& opt)
         pipeline_softmax_div_sum_pack4->create("softmax_div_sum_pack4", opt, specializations, 2, 10);
     }
 
+    // pack8
+    {
+        pipeline_softmax_reduce_max_pack8 = new Pipeline(vkdev);
+        pipeline_softmax_exp_sub_max_pack8 = new Pipeline(vkdev);
+        pipeline_softmax_reduce_sum_pack8 = new Pipeline(vkdev);
+        pipeline_softmax_div_sum_pack8 = new Pipeline(vkdev);
+
+        pipeline_softmax_reduce_max_pack8->set_optimal_local_size_xyz();
+        pipeline_softmax_exp_sub_max_pack8->set_optimal_local_size_xyz();
+        pipeline_softmax_reduce_sum_pack8->set_optimal_local_size_xyz();
+        pipeline_softmax_div_sum_pack8->set_optimal_local_size_xyz();
+
+        pipeline_softmax_reduce_max_pack8->create("softmax_reduce_max_pack8", opt, specializations, 2, 10);
+        pipeline_softmax_exp_sub_max_pack8->create("softmax_exp_sub_max_pack8", opt, specializations, 2, 10);
+        pipeline_softmax_reduce_sum_pack8->create("softmax_reduce_sum_pack8", opt, specializations, 2, 10);
+        pipeline_softmax_div_sum_pack8->create("softmax_div_sum_pack8", opt, specializations, 2, 10);
+    }
+
     return 0;
 }
 
@@ -106,6 +129,18 @@ int Softmax_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_softmax_div_sum_pack4;
     pipeline_softmax_div_sum_pack4 = 0;
 
+    delete pipeline_softmax_reduce_max_pack8;
+    pipeline_softmax_reduce_max_pack8 = 0;
+
+    delete pipeline_softmax_exp_sub_max_pack8;
+    pipeline_softmax_exp_sub_max_pack8 = 0;
+
+    delete pipeline_softmax_reduce_sum_pack8;
+    pipeline_softmax_reduce_sum_pack8 = 0;
+
+    delete pipeline_softmax_div_sum_pack8;
+    pipeline_softmax_div_sum_pack8 = 0;
+
     return 0;
 }
 
@@ -170,7 +205,9 @@ int Softmax_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, cons
     constants[8].i = max_workspace.c;
     constants[9].i = max_workspace.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_softmax_reduce_max_pack4 : pipeline_softmax_reduce_max;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_softmax_reduce_max_pack8
+                             : elempack == 4 ? pipeline_softmax_reduce_max_pack4
+                             : pipeline_softmax_reduce_max;
 
     cmd.record_pipeline(pipeline, bindings, constants, max_workspace);
     }
@@ -193,7 +230,9 @@ int Softmax_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, cons
     constants[8].i = max_workspace.c;
     constants[9].i = max_workspace.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_softmax_exp_sub_max_pack4 : pipeline_softmax_exp_sub_max;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_softmax_exp_sub_max_pack8
+                             : elempack == 4 ? pipeline_softmax_exp_sub_max_pack4
+                             : pipeline_softmax_exp_sub_max;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
     }
@@ -216,7 +255,9 @@ int Softmax_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, cons
     constants[8].i = sum_workspace.c;
     constants[9].i = sum_workspace.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_softmax_reduce_sum_pack4 : pipeline_softmax_reduce_sum;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_softmax_reduce_sum_pack8
+                             : elempack == 4 ? pipeline_softmax_reduce_sum_pack4
+                             : pipeline_softmax_reduce_sum;
 
     cmd.record_pipeline(pipeline, bindings, constants, sum_workspace);
     }
@@ -239,7 +280,9 @@ int Softmax_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, cons
     constants[8].i = sum_workspace.c;
     constants[9].i = sum_workspace.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_softmax_div_sum_pack4 : pipeline_softmax_div_sum;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_softmax_div_sum_pack8
+                             : elempack == 4 ? pipeline_softmax_div_sum_pack4
+                             : pipeline_softmax_div_sum;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
     }

src/layer/vulkan/softmax_vulkan.h

@@ -40,6 +40,11 @@ public:
     Pipeline* pipeline_softmax_exp_sub_max_pack4;
     Pipeline* pipeline_softmax_reduce_sum_pack4;
     Pipeline* pipeline_softmax_div_sum_pack4;
+
+    Pipeline* pipeline_softmax_reduce_max_pack8;
+    Pipeline* pipeline_softmax_exp_sub_max_pack8;
+    Pipeline* pipeline_softmax_reduce_sum_pack8;
+    Pipeline* pipeline_softmax_div_sum_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/tanh_vulkan.cpp

@@ -25,6 +25,7 @@ TanH_vulkan::TanH_vulkan()
 
     pipeline_tanh = 0;
     pipeline_tanh_pack4 = 0;
+    pipeline_tanh_pack8 = 0;
 }
 
 int TanH_vulkan::create_pipeline(const Option& opt)
@@ -45,6 +46,13 @@ int TanH_vulkan::create_pipeline(const Option& opt)
         pipeline_tanh_pack4->create("tanh_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_tanh_pack8 = new Pipeline(vkdev);
+        pipeline_tanh_pack8->set_optimal_local_size_xyz();
+        pipeline_tanh_pack8->create("tanh_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -56,6 +64,9 @@ int TanH_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_tanh_pack4;
     pipeline_tanh_pack4 = 0;
 
+    delete pipeline_tanh_pack8;
+    pipeline_tanh_pack8 = 0;
+
     return 0;
 }
 
@@ -73,7 +84,9 @@ int TanH_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const O
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_tanh_pack4 : pipeline_tanh;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_tanh_pack8
+                             : elempack == 4 ? pipeline_tanh_pack4
+                             : pipeline_tanh;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/tanh_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_tanh;
     Pipeline* pipeline_tanh_pack4;
+    Pipeline* pipeline_tanh_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/unaryop_vulkan.cpp

@@ -24,6 +24,7 @@ UnaryOp_vulkan::UnaryOp_vulkan()
 
     pipeline_unaryop = 0;
     pipeline_unaryop_pack4 = 0;
+    pipeline_unaryop_pack8 = 0;
 }
 
 int UnaryOp_vulkan::create_pipeline(const Option& opt)
@@ -45,6 +46,13 @@ int UnaryOp_vulkan::create_pipeline(const Option& opt)
         pipeline_unaryop_pack4->create("unaryop_pack4", opt, specializations, 1, 5);
     }
 
+    // pack8
+    {
+        pipeline_unaryop_pack8 = new Pipeline(vkdev);
+        pipeline_unaryop_pack8->set_optimal_local_size_xyz();
+        pipeline_unaryop_pack8->create("unaryop_pack8", opt, specializations, 1, 5);
+    }
+
     return 0;
 }
 
@@ -56,6 +64,9 @@ int UnaryOp_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_unaryop_pack4;
     pipeline_unaryop_pack4 = 0;
 
+    delete pipeline_unaryop_pack8;
+    pipeline_unaryop_pack8 = 0;
+
     return 0;
 }
 
@@ -73,7 +84,9 @@ int UnaryOp_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, cons
     constants[3].i = bottom_top_blob.c;
     constants[4].i = bottom_top_blob.cstep;
 
-    const Pipeline* pipeline = elempack == 4 ? pipeline_unaryop_pack4 : pipeline_unaryop;
+    const Pipeline* pipeline = elempack == 8 ? pipeline_unaryop_pack8
+                             : elempack == 4 ? pipeline_unaryop_pack4
+                             : pipeline_unaryop;
 
     cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
 

src/layer/vulkan/unaryop_vulkan.h

@@ -33,6 +33,7 @@ public:
 public:
     Pipeline* pipeline_unaryop;
     Pipeline* pipeline_unaryop_pack4;
+    Pipeline* pipeline_unaryop_pack8;
 };
 
 } // namespace ncnn