 [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  new int8 implement,better accuracy (#749)
* add the armv7a conv3x3s1 implement without overflow,remove old codes
* fix the bug of conv3x3s2 packed int8
* new int8 implement,weight quant by perchanel,better accuracy~
* fix the bug of conv3x3s1 packed int8 neon
* add the naive c fp32 and int8 winograd F(2,3)
* add the neon intrinsic int8 winograd F(2,3)
* optimize the armv7a int8 winograd F(2,3) with neon assembly
* optimize the armv7a int8 winograd F(2,3) input transform with assembly.
* add the requantize layer and int8 relu implement.
* add graph optimize conv1x1s2 -> conv1x1s1,begin optimize int8 aarch64.
* fix int8 bugs
* add the c naive im2col with sgemm
* add aarch64 int8 winograd f23, conv3x3s2 naive implement
* add the int8 sgemm conv7x7s2 on x86/armv7a platform
* optimize the int8 sgemm by neon intrinsic and packed kernel
* optimize the int8 sgemm with packed data
* optimize the int8 sgemm with armv7a neon assembly
* add the int8 sgemm on arm64-v8a platform
* perpare to merge latest codes from master
* add the int8 param files
* In the Class Net,add the fuse_network method
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago  [WIP] vulkan compute (#618)
* vulkan infrastructure
* vkallocator and vkmat
* layer interface for vulkan compute
* wip...
* default vulkan device, command wrapper, upload model weight in load_model to simplify layer interface
* simplify command api, vkmat holds staging buffer, relu works
* initialize specialization constant, simplify command dispatch, fix staging buffer copy with different shape, convolution works
* init extension functions
* dynamic local size and group count
* group count=1 is invalid
* regard device max workgroup size limit
* fix relu oooops
* decouple command record and staging allocation
* create result blob
* add pooling shader
* buffer is faster than image :)
* fix pooling shader
* add innerproduct shader
* readonly writeonly decoration
* simplify buffer creation
* decouple command and layer, VK_KHR_descriptor_update_template extension makes descriptor binding update easy :D
* fix vulkan building issues in visual studio (#1)
* fix building issues on visual studio
* ignore benchmark
* cancel changes
* ... ...
* decouple paramdict and vulkandevice
* fix staging buffer destroy in model loading
* remove vkdev member in option
* add padding shader
* simplify vulkan layer creation, simplify convolution and pooling shader for no padding, less debug output
* add convolutiondepthwise and softmax shader
* specialization float type, add leakyrelu
* add dropout shader
* add batchnorm shader
* split vulkan forward
* add scale shader
* push constant type can be int or float
* set_optimal_local_size_xyz
* add eltwise shader
* concat vulkan forward
* fix convolution without bias
* add dummy shader for concat and split, more fix ...
* optional VK_KHR_descriptor_update_template and VK_KHR_push_descriptor
* check VK_KHR_push_descriptor for vkCmdPushDescriptorSetWithTemplateKHR
* binaryop and unaryop shader
* hide raw command buffer
* simple vkbenchncnn benchmark
* create device with transfer queue
* rename command to vkcompute, add vktransfer and layer upload_model interface
* external VkMat, copy and map wrt buffer offset
* command copy respect offset and size
* decouple weight upload and load, simplify upload weight api, use one big staging buffer for uploading weights
* fix build on android
* binding count can not vary :(
* barrier check state, fix sub-op destruction
* declare local_size_xyz constant, fix crash on radv
* fix local_size_xyz, second try
* more barrier and state fix
* fix softmax
* reconstruct buffer memory allocator, reuse blob buffer, less verbose output
* find unified memory type index
* weight staging buffer allocator and weight buffer allocator, respect descriptor buffer offset alignment
* use VK_KHR_descriptor_update_template for faster descriptor update if available, multithread pipeline creation
* find more useful vulkan extensions and enable them
* fix msvc build
* respect VK_KHR_dedicated_allocation for weight buffer allocation
* fix android build
* fix bias name conflicts with metal
* decouple pipeline and layer, building shader sources into shader module, dedicated create_pipeline api, simplify pipeline recording
* drop dummy shader, inplace softmax, multiple shader module works
* fix unique queue family index error
* flatten support vulkan
* mnasnet run
* find shader module by name, each entry point per shader module, fix attribute/id conflict on moltenvk
* some minor changes
* add some high level api
* use dedicated transfer queue to upload weight model
* prefer mappable buffer on unified memory
* global pooling and convolution fc, reuse staging buffer
* implement ring-buffer style blob allocator, add VkBufferMemory capacity
* use blob allocator for workspace blob, it works fine :)
* vulkan option off
* Update layer.cpp
* fix build with vulkan off
* less verbose output, fix crash on vulkan_compute off
* merge benchncnn tool
* allocator clear api, use new weight buffer allocator per net
* add default locked allocator
* mapped mat ptr api, persistent mapped memory works generally :)
* travis ci linux vulkan
* travis ci vulkan wip ...
* more gpu wip ...
* more gpu wip ...
* wip...
* wip...
* wip... ...
* wip... ios vulkan build...
* find glslangValidator on ios build
* use dynamic moltenvk library
* travis ci wip ...
* ios simulator does not support metal at all
* fix cpu only extractor
* optimize workgroup size, first try
* optimize workgroup size, second try
* conv1x1s1d1 vec4
* revert build system
* fix ncnn2mem build
* fix ncnn2mem build
7 years ago |
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807 |
- // Tencent is pleased to support the open source community by making ncnn available.
- //
- // Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
- //
- // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
- // in compliance with the License. You may obtain a copy of the License at
- //
- // https://opensource.org/licenses/BSD-3-Clause
- //
- // Unless required by applicable law or agreed to in writing, software distributed
- // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
- // CONDITIONS OF ANY KIND, either express or implied. See the License for the
- // specific language governing permissions and limitations under the License.
-
- #include "net.h"
-
- #include "cpu.h"
- #include "datareader.h"
- #include "layer_type.h"
- #include "modelbin.h"
- #include "paramdict.h"
-
- #include <stdarg.h>
- #include <stdint.h>
- #include <string.h>
-
- #if NCNN_BENCHMARK
- #include "benchmark.h"
- #endif // NCNN_BENCHMARK
-
- #if NCNN_VULKAN
- #include "command.h"
- #include "pipelinecache.h"
- #endif // NCNN_VULKAN
-
- namespace ncnn {
-
- class NetPrivate
- {
- public:
- NetPrivate(Option& _opt);
-
- Option& opt;
-
- #if NCNN_VULKAN
-
- int upload_model();
-
- #endif // NCNN_VULKAN
-
- friend class Extractor;
- int forward_layer(int layer_index, std::vector<Mat>& blob_mats, const Option& opt) const;
-
- #if NCNN_VULKAN
- int forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector<VkMat>& blob_mats_gpu, VkCompute& cmd, const Option& opt) const;
- int forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector<VkMat>& blob_mats_gpu, std::vector<VkImageMat>& blob_mats_gpu_image, VkCompute& cmd, const Option& opt) const;
- #endif // NCNN_VULKAN
-
- int convert_layout(Mat& bottom_blob, const Layer* layer, const Option& opt) const;
-
- int do_forward_layer(const Layer* layer, std::vector<Mat>& blob_mats, const Option& opt) const;
- #if NCNN_VULKAN
- int do_forward_layer(const Layer* layer, std::vector<VkMat>& blob_mats_gpu, VkCompute& cmd, const Option& opt) const;
- int do_forward_layer(const Layer* layer, std::vector<VkImageMat>& blob_mats_gpu_image, VkCompute& cmd, const Option& opt) const;
- #endif // NCNN_VULKAN
-
- void update_input_output_indexes();
- #if NCNN_STRING
- void update_input_output_names();
- #endif // NCNN_STRING
-
- std::vector<Blob> blobs;
- std::vector<Layer*> layers;
-
- std::vector<int> input_blob_indexes;
- std::vector<int> output_blob_indexes;
- #if NCNN_STRING
- std::vector<const char*> input_blob_names;
- std::vector<const char*> output_blob_names;
- #endif // NCNN_STRING
-
- std::vector<custom_layer_registry_entry> custom_layer_registry;
-
- PoolAllocator* local_blob_allocator;
- PoolAllocator* local_workspace_allocator;
-
- #if NCNN_VULKAN
- const VulkanDevice* vkdev;
-
- VkAllocator* weight_vkallocator;
- VkAllocator* weight_staging_vkallocator;
-
- PipelineCache* pipeline_cache;
- #endif // NCNN_VULKAN
- };
-
- NetPrivate::NetPrivate(Option& _opt)
- : opt(_opt)
- {
- local_blob_allocator = 0;
- local_workspace_allocator = 0;
-
- #if NCNN_VULKAN
- vkdev = 0;
- weight_vkallocator = 0;
- weight_staging_vkallocator = 0;
- pipeline_cache = 0;
- #endif // NCNN_VULKAN
- }
-
- #if NCNN_VULKAN
- int NetPrivate::upload_model()
- {
- ncnn::VkTransfer cmd(vkdev);
-
- // create gpu device allocator if null
- if (!weight_vkallocator)
- {
- weight_vkallocator = new VkWeightAllocator(vkdev);
- }
- if (!weight_staging_vkallocator)
- {
- weight_staging_vkallocator = new VkWeightStagingAllocator(vkdev);
- }
-
- Option opt_upload = opt;
- opt_upload.blob_vkallocator = weight_vkallocator;
- opt_upload.workspace_vkallocator = weight_vkallocator;
- opt_upload.staging_vkallocator = weight_staging_vkallocator;
-
- for (size_t i = 0; i < layers.size(); i++)
- {
- if (layers[i]->support_vulkan)
- {
- int uret = layers[i]->upload_model(cmd, opt_upload);
- if (uret != 0)
- {
- NCNN_LOGE("layer upload_model %d failed", (int)i);
- return -1;
- }
- }
- }
-
- cmd.submit_and_wait();
-
- return 0;
- }
- #endif // NCNN_VULKAN
-
- int NetPrivate::forward_layer(int layer_index, std::vector<Mat>& blob_mats, const Option& opt) const
- {
- const Layer* layer = layers[layer_index];
-
- // NCNN_LOGE("forward_layer %d %s", layer_index, layer->name.c_str());
-
- if (layer->one_blob_only)
- {
- // load bottom blob
- int bottom_blob_index = layer->bottoms[0];
-
- if (blob_mats[bottom_blob_index].dims == 0)
- {
- int ret = forward_layer(blobs[bottom_blob_index].producer, blob_mats, opt);
- if (ret != 0)
- return ret;
- }
- }
- else
- {
- // load bottom blobs
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- if (blob_mats[bottom_blob_index].dims == 0)
- {
- int ret = forward_layer(blobs[bottom_blob_index].producer, blob_mats, opt);
- if (ret != 0)
- return ret;
- }
- }
- }
-
- #if NCNN_BENCHMARK
- double start = get_current_time();
- Mat bottom_blob;
- if (layer->one_blob_only)
- {
- int bottom_blob_index = layer->bottoms[0];
- bottom_blob.dims = blob_mats[bottom_blob_index].dims;
- bottom_blob.w = blob_mats[bottom_blob_index].w;
- bottom_blob.h = blob_mats[bottom_blob_index].h;
- bottom_blob.c = blob_mats[bottom_blob_index].c;
- bottom_blob.elempack = blob_mats[bottom_blob_index].elempack;
- bottom_blob.elemsize = blob_mats[bottom_blob_index].elemsize;
- }
- #endif
- int ret = do_forward_layer(layer, blob_mats, opt);
- #if NCNN_BENCHMARK
- double end = get_current_time();
- if (layer->one_blob_only)
- {
- int top_blob_index = layer->tops[0];
- benchmark(layer, bottom_blob, blob_mats[top_blob_index], start, end);
- }
- else
- {
- benchmark(layer, start, end);
- }
- #endif
- if (ret != 0)
- return ret;
-
- // NCNN_LOGE("forward_layer %d %s done", layer_index, layer->name.c_str());
- // const Mat& blob = blob_mats[layer->tops[0]];
- // NCNN_LOGE("[%-2d %-16s %-16s] %d blobs count = %-3d size = %-3d x %-3d", layer_index, layer->type.c_str(), layer->name.c_str(), layer->tops[0], blob.c, blob.h, blob.w);
-
- return 0;
- }
-
- #if NCNN_VULKAN
- int NetPrivate::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector<VkMat>& blob_mats_gpu, VkCompute& cmd, const Option& opt) const
- {
- const Layer* layer = layers[layer_index];
-
- // NCNN_LOGE("forward_layer %d %d %s", layer->support_vulkan, layer_index, layer->name.c_str());
-
- bool cmd_submit_and_wait = false;
-
- if (layer->one_blob_only)
- {
- // load bottom blob
- int bottom_blob_index = layer->bottoms[0];
-
- if (blob_mats_gpu[bottom_blob_index].dims == 0 && blob_mats[bottom_blob_index].dims == 0)
- {
- int ret = forward_layer(blobs[bottom_blob_index].producer, blob_mats, blob_mats_gpu, cmd, opt);
- if (ret != 0)
- return ret;
- }
-
- if (layer->support_vulkan)
- {
- if (blob_mats_gpu[bottom_blob_index].dims == 0)
- {
- // host to buffer
- cmd.record_upload(blob_mats[bottom_blob_index], blob_mats_gpu[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats[bottom_blob_index].release();
- }
- }
- }
- else
- {
- if (blob_mats[bottom_blob_index].dims == 0)
- {
- Option opt_download = opt;
- opt_download.use_packing_layout = layer->support_packing;
-
- // buffer to host
- cmd.record_download(blob_mats_gpu[bottom_blob_index], blob_mats[bottom_blob_index], opt_download);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu[bottom_blob_index].release();
- }
-
- cmd_submit_and_wait = true;
- }
- }
- }
- else
- {
- // load bottom blobs
- std::vector<VkMat> bottom_blobs(layer->bottoms.size());
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- if (blob_mats_gpu[bottom_blob_index].dims == 0 && blob_mats[bottom_blob_index].dims == 0)
- {
- int ret = forward_layer(blobs[bottom_blob_index].producer, blob_mats, blob_mats_gpu, cmd, opt);
- if (ret != 0)
- return ret;
- }
-
- if (layer->support_vulkan)
- {
- if (blob_mats_gpu[bottom_blob_index].dims == 0)
- {
- // host to buffer
- cmd.record_upload(blob_mats[bottom_blob_index], blob_mats_gpu[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats[bottom_blob_index].release();
- }
- }
- }
- else
- {
- if (blob_mats[bottom_blob_index].dims == 0)
- {
- Option opt_download = opt;
- opt_download.use_packing_layout = layer->support_packing;
-
- // buffer to host
- cmd.record_download(blob_mats_gpu[bottom_blob_index], blob_mats[bottom_blob_index], opt_download);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu[bottom_blob_index].release();
- }
-
- cmd_submit_and_wait = true;
- }
- }
- }
- }
-
- if (cmd_submit_and_wait)
- {
- cmd.submit_and_wait();
-
- #if NCNN_BENCHMARK
- std::vector<uint64_t> results(layer_index * 2);
- cmd.get_query_pool_results(0, layer_index * 2, results);
- for (int i = 0; i < layer_index; i++)
- {
- uint64_t start = results[i * 2];
- uint64_t end = results[i * 2 + 1];
- if (start == 0 || end == 0)
- continue;
-
- double duration_us = (end - start) * vkdev->info.timestamp_period() / 1000;
- NCNN_LOGE("%-24s %-30s %8.2lfus |", layers[i]->type.c_str(), layers[i]->name.c_str(), duration_us);
- }
- #endif // NCNN_BENCHMARK
-
- cmd.reset();
- }
-
- int ret;
- if (layer->support_vulkan)
- {
- #if NCNN_BENCHMARK
- cmd.record_write_timestamp(layer_index * 2);
- #endif
- ret = do_forward_layer(layer, blob_mats_gpu, cmd, opt);
- #if NCNN_BENCHMARK
- cmd.record_write_timestamp(layer_index * 2 + 1);
- #endif
- }
- else
- {
- #if NCNN_BENCHMARK
- double start = get_current_time();
- Mat bottom_blob;
- if (layer->one_blob_only)
- {
- int bottom_blob_index = layer->bottoms[0];
- bottom_blob = blob_mats[bottom_blob_index].shape();
- }
- #endif
- ret = do_forward_layer(layer, blob_mats, opt);
- #if NCNN_BENCHMARK
- double end = get_current_time();
- if (layer->one_blob_only)
- {
- int top_blob_index = layer->tops[0];
- benchmark(layer, bottom_blob, blob_mats[top_blob_index], start, end);
- }
- else
- {
- benchmark(layer, start, end);
- }
- #endif
- }
- if (ret != 0)
- return ret;
-
- // NCNN_LOGE("forward_layer %d %d %s done", layer->support_vulkan, layer_index, layer->name.c_str());
-
- return 0;
- }
-
- int NetPrivate::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector<VkMat>& blob_mats_gpu, std::vector<VkImageMat>& blob_mats_gpu_image, VkCompute& cmd, const Option& opt) const
- {
- const Layer* layer = layers[layer_index];
-
- // NCNN_LOGE("forward_layer %d %d %s", layer->support_vulkan, layer_index, layer->name.c_str());
-
- bool cmd_submit_and_wait = false;
- bool image_allocation_failed = false;
-
- IMAGE_ALLOCATION_FAILED:
-
- if (image_allocation_failed)
- {
- #if NCNN_STRING
- NCNN_LOGE("forward_layer %d %s image allocation failed, fallback to cpu", layer_index, layer->name.c_str());
- #else
- NCNN_LOGE("forward_layer %d image allocation failed, fallback to cpu", layer_index);
- #endif
- }
-
- if (layer->one_blob_only)
- {
- // load bottom blob
- int bottom_blob_index = layer->bottoms[0];
-
- if (blob_mats_gpu_image[bottom_blob_index].dims == 0 && blob_mats_gpu[bottom_blob_index].dims == 0 && blob_mats[bottom_blob_index].dims == 0)
- {
- int ret = forward_layer(blobs[bottom_blob_index].producer, blob_mats, blob_mats_gpu, blob_mats_gpu_image, cmd, opt);
- if (ret != 0)
- return ret;
- }
-
- if (layer->support_vulkan && !image_allocation_failed)
- {
- if (layer->support_image_storage)
- {
- if (blob_mats_gpu_image[bottom_blob_index].dims == 0)
- {
- if (blob_mats_gpu[bottom_blob_index].dims == 0)
- {
- // host to image
- cmd.record_upload(blob_mats[bottom_blob_index], blob_mats_gpu_image[bottom_blob_index], opt);
-
- if (blob_mats_gpu_image[bottom_blob_index].empty())
- {
- image_allocation_failed = true;
- goto IMAGE_ALLOCATION_FAILED;
- }
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats[bottom_blob_index].release();
- }
- }
- else
- {
- // buffer to image
- cmd.record_buffer_to_image(blob_mats_gpu[bottom_blob_index], blob_mats_gpu_image[bottom_blob_index], opt);
-
- if (blob_mats_gpu_image[bottom_blob_index].empty())
- {
- image_allocation_failed = true;
- goto IMAGE_ALLOCATION_FAILED;
- }
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu[bottom_blob_index].release();
- }
- }
- }
- }
- else
- {
- if (blob_mats_gpu[bottom_blob_index].dims == 0)
- {
- if (blob_mats_gpu_image[bottom_blob_index].dims == 0)
- {
- // host to buffer
- cmd.record_upload(blob_mats[bottom_blob_index], blob_mats_gpu[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats[bottom_blob_index].release();
- }
- }
- else
- {
- // image to buffer
- cmd.record_image_to_buffer(blob_mats_gpu_image[bottom_blob_index], blob_mats_gpu[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu_image[bottom_blob_index].release();
- }
- }
- }
- }
- }
- else
- {
- if (blob_mats[bottom_blob_index].dims == 0)
- {
- if (blob_mats_gpu_image[bottom_blob_index].dims == 0)
- {
- // buffer to host
- cmd.record_download(blob_mats_gpu[bottom_blob_index], blob_mats[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu[bottom_blob_index].release();
- }
-
- cmd_submit_and_wait = true;
- }
- else
- {
- // image to host
- cmd.record_download(blob_mats_gpu_image[bottom_blob_index], blob_mats[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu_image[bottom_blob_index].release();
- }
-
- cmd_submit_and_wait = true;
- }
- }
- }
- }
- else
- {
- // load bottom blobs
- std::vector<VkImageMat> bottom_blobs(layer->bottoms.size());
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- if (blob_mats_gpu_image[bottom_blob_index].dims == 0 && blob_mats_gpu[bottom_blob_index].dims == 0 && blob_mats[bottom_blob_index].dims == 0)
- {
- int ret = forward_layer(blobs[bottom_blob_index].producer, blob_mats, blob_mats_gpu, blob_mats_gpu_image, cmd, opt);
- if (ret != 0)
- return ret;
- }
-
- if (layer->support_vulkan && !image_allocation_failed)
- {
- if (layer->support_image_storage)
- {
- if (blob_mats_gpu_image[bottom_blob_index].dims == 0)
- {
- if (blob_mats_gpu[bottom_blob_index].dims == 0)
- {
- // host to image
- cmd.record_upload(blob_mats[bottom_blob_index], blob_mats_gpu_image[bottom_blob_index], opt);
-
- if (blob_mats_gpu_image[bottom_blob_index].empty())
- {
- image_allocation_failed = true;
- goto IMAGE_ALLOCATION_FAILED;
- }
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats[bottom_blob_index].release();
- }
- }
- else
- {
- // buffer to image
- cmd.record_buffer_to_image(blob_mats_gpu[bottom_blob_index], blob_mats_gpu_image[bottom_blob_index], opt);
-
- if (blob_mats_gpu_image[bottom_blob_index].empty())
- {
- image_allocation_failed = true;
- goto IMAGE_ALLOCATION_FAILED;
- }
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu[bottom_blob_index].release();
- }
- }
- }
- }
- else
- {
- if (blob_mats_gpu[bottom_blob_index].dims == 0)
- {
- if (blob_mats_gpu_image[bottom_blob_index].dims == 0)
- {
- // host to buffer
- cmd.record_upload(blob_mats[bottom_blob_index], blob_mats_gpu[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats[bottom_blob_index].release();
- }
- }
- else
- {
- // image to buffer
- cmd.record_image_to_buffer(blob_mats_gpu_image[bottom_blob_index], blob_mats_gpu[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu_image[bottom_blob_index].release();
- }
- }
- }
- }
- }
- else
- {
- if (blob_mats[bottom_blob_index].dims == 0)
- {
- if (blob_mats_gpu_image[bottom_blob_index].dims == 0)
- {
- // buffer to host
- cmd.record_download(blob_mats_gpu[bottom_blob_index], blob_mats[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu[bottom_blob_index].release();
- }
-
- cmd_submit_and_wait = true;
- }
- else
- {
- // image to host
- cmd.record_download(blob_mats_gpu_image[bottom_blob_index], blob_mats[bottom_blob_index], opt);
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu_image[bottom_blob_index].release();
- }
-
- cmd_submit_and_wait = true;
- }
- }
- }
- }
- }
-
- if (cmd_submit_and_wait)
- {
- cmd.submit_and_wait();
-
- #if NCNN_BENCHMARK
- std::vector<uint64_t> results(layer_index * 2);
- cmd.get_query_pool_results(0, layer_index * 2, results);
- for (int i = 0; i < layer_index; i++)
- {
- uint64_t start = results[i * 2];
- uint64_t end = results[i * 2 + 1];
- if (start == 0 || end == 0)
- continue;
-
- double duration_us = (end - start) * vkdev->info.timestamp_period() / 1000;
- NCNN_LOGE("%-24s %-30s %8.2lfus |", layers[i]->type.c_str(), layers[i]->name.c_str(), duration_us);
- }
- #endif // NCNN_BENCHMARK
-
- cmd.reset();
- }
-
- int ret;
- if (layer->support_vulkan && !image_allocation_failed)
- {
- #if NCNN_BENCHMARK
- cmd.record_write_timestamp(layer_index * 2);
- #endif
- if (layer->support_image_storage)
- {
- ret = do_forward_layer(layer, blob_mats_gpu_image, cmd, opt);
- if (ret == -100)
- {
- image_allocation_failed = true;
- goto IMAGE_ALLOCATION_FAILED;
- }
- }
- else
- {
- ret = do_forward_layer(layer, blob_mats_gpu, cmd, opt);
- }
- #if NCNN_BENCHMARK
- cmd.record_write_timestamp(layer_index * 2 + 1);
- #endif
- }
- else
- {
- #if NCNN_BENCHMARK
- double start = get_current_time();
- Mat bottom_blob;
- if (layer->one_blob_only)
- {
- int bottom_blob_index = layer->bottoms[0];
- bottom_blob = blob_mats[bottom_blob_index].shape();
- }
- #endif
- ret = do_forward_layer(layer, blob_mats, opt);
- #if NCNN_BENCHMARK
- double end = get_current_time();
- if (layer->one_blob_only)
- {
- int top_blob_index = layer->tops[0];
- benchmark(layer, bottom_blob, blob_mats[top_blob_index], start, end);
- }
- else
- {
- benchmark(layer, start, end);
- }
- #endif
- }
- if (ret != 0)
- return ret;
-
- // NCNN_LOGE("forward_layer %d %d %s done", layer->support_vulkan, layer_index, layer->name.c_str());
-
- return 0;
- }
- #endif // NCNN_VULKAN
-
- int NetPrivate::convert_layout(Mat& bottom_blob, const Layer* layer, const Option& opt) const
- {
- // clang-format off
- // *INDENT-OFF*
- #if NCNN_ARM82
- if (opt.use_fp16_storage && cpu_support_arm_asimdhp())
- {
- if (bottom_blob.elembits() == 32 && layer->support_fp16_storage)
- {
- Mat bottom_blob_fp16;
- cast_float32_to_float16(bottom_blob, bottom_blob_fp16, opt);
- bottom_blob = bottom_blob_fp16;
- }
- if (bottom_blob.elembits() == 16 && !layer->support_fp16_storage)
- {
- Mat bottom_blob_fp32;
- cast_float16_to_float32(bottom_blob, bottom_blob_fp32, opt);
- bottom_blob = bottom_blob_fp32;
- }
- }
- else
- #endif // NCNN_ARM82
- #if NCNN_RVV
- if (opt.use_fp16_storage && cpu_support_riscv_v() && cpu_support_riscv_zfh())
- {
- if (bottom_blob.elembits() == 32 && layer->support_fp16_storage)
- {
- Mat bottom_blob_fp16;
- cast_float32_to_float16(bottom_blob, bottom_blob_fp16, opt);
- bottom_blob = bottom_blob_fp16;
- }
- if (bottom_blob.elembits() == 16 && !layer->support_fp16_storage)
- {
- Mat bottom_blob_fp32;
- cast_float16_to_float32(bottom_blob, bottom_blob_fp32, opt);
- bottom_blob = bottom_blob_fp32;
- }
- }
- else
- #endif // NCNN_RVV
- #if NCNN_BF16
- if (opt.use_bf16_storage)
- {
- if (bottom_blob.elembits() == 32 && layer->support_bf16_storage)
- {
- Mat bottom_blob_bf16;
- cast_float32_to_bfloat16(bottom_blob, bottom_blob_bf16, opt);
- bottom_blob = bottom_blob_bf16;
- }
- if (bottom_blob.elembits() == 16 && !layer->support_bf16_storage)
- {
- Mat bottom_blob_fp32;
- cast_bfloat16_to_float32(bottom_blob, bottom_blob_fp32, opt);
- bottom_blob = bottom_blob_fp32;
- }
- }
- else
- #endif // NCNN_BF16
- {
- // no type conversion
- }
- // *INDENT-ON*
- // clang-format on
-
- if (opt.use_packing_layout)
- {
- // resolve dst_elempack
- int dims = bottom_blob.dims;
- int elemcount = 0;
- if (dims == 1) elemcount = bottom_blob.elempack * bottom_blob.w;
- if (dims == 2) elemcount = bottom_blob.elempack * bottom_blob.h;
- if (dims == 3) elemcount = bottom_blob.elempack * bottom_blob.c;
-
- int elembits = bottom_blob.elembits();
-
- int dst_elempack = 1;
- if (layer->support_packing)
- {
- if (elembits == 32)
- {
- #if (NCNN_AVX2 || NCNN_AVX)
- if (elemcount % 8 == 0 && (ncnn::cpu_support_x86_avx2() || ncnn::cpu_support_x86_avx()))
- dst_elempack = 8;
- else if (elemcount % 4 == 0)
- dst_elempack = 4;
- #elif NCNN_RVV
- const int packn = ncnn::cpu_riscv_vlenb() / 4;
- if (elemcount % packn == 0)
- dst_elempack = packn;
- #else
- if (elemcount % 4 == 0)
- dst_elempack = 4;
- #endif
- }
- if (elembits == 16)
- {
- #if NCNN_ARM82
- if (elemcount % 8 == 0 && opt.use_fp16_storage && opt.use_fp16_arithmetic && layer->support_fp16_storage)
- dst_elempack = 8;
- else if (elemcount % 4 == 0)
- dst_elempack = 4;
- #elif NCNN_RVV
- const int packn = ncnn::cpu_riscv_vlenb() / 2;
- if (elemcount % packn == 0)
- dst_elempack = packn;
- #else
- if (elemcount % 4 == 0)
- dst_elempack = 4;
- #endif
- }
- if (elembits == 8)
- {
- #if NCNN_RVV
- const int packn = ncnn::cpu_riscv_vlenb() / 1;
- if (elemcount % packn == 0)
- dst_elempack = packn;
- #else
- if (elemcount % 8 == 0)
- dst_elempack = 8;
- #endif
- }
- }
-
- if (bottom_blob.elempack != dst_elempack)
- {
- Mat bottom_blob_packed;
- convert_packing(bottom_blob, bottom_blob_packed, dst_elempack, opt);
- bottom_blob = bottom_blob_packed;
- }
- }
-
- return 0;
- }
-
- int NetPrivate::do_forward_layer(const Layer* layer, std::vector<Mat>& blob_mats, const Option& opt) const
- {
- if (layer->one_blob_only)
- {
- int bottom_blob_index = layer->bottoms[0];
- int top_blob_index = layer->tops[0];
-
- Mat& bottom_blob_ref = blob_mats[bottom_blob_index];
- Mat bottom_blob;
-
- if (opt.lightmode)
- {
- // deep copy for inplace forward if data is shared
- if (layer->support_inplace && *bottom_blob_ref.refcount != 1)
- {
- bottom_blob = bottom_blob_ref.clone(opt.blob_allocator);
- }
- }
- if (bottom_blob.dims == 0)
- {
- bottom_blob = bottom_blob_ref;
- }
-
- convert_layout(bottom_blob, layer, opt);
-
- // forward
- if (opt.lightmode && layer->support_inplace)
- {
- Mat& bottom_top_blob = bottom_blob;
- int ret = layer->forward_inplace(bottom_top_blob, opt);
- if (ret != 0)
- return ret;
-
- // store top blob
- blob_mats[top_blob_index] = bottom_top_blob;
- }
- else
- {
- Mat top_blob;
- int ret = layer->forward(bottom_blob, top_blob, opt);
- if (ret != 0)
- return ret;
-
- // store top blob
- blob_mats[top_blob_index] = top_blob;
- }
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats[bottom_blob_index].release();
- }
- }
- else
- {
- std::vector<Mat> bottom_blobs(layer->bottoms.size());
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- Mat& bottom_blob_ref = blob_mats[bottom_blob_index];
- bottom_blobs[i].release();
-
- if (opt.lightmode)
- {
- // deep copy for inplace forward if data is shared
- if (layer->support_inplace && *bottom_blob_ref.refcount != 1)
- {
- bottom_blobs[i] = bottom_blob_ref.clone(opt.blob_allocator);
- }
- }
- if (bottom_blobs[i].dims == 0)
- {
- bottom_blobs[i] = bottom_blob_ref;
- }
-
- convert_layout(bottom_blobs[i], layer, opt);
- }
-
- // forward
- if (opt.lightmode && layer->support_inplace)
- {
- std::vector<Mat>& bottom_top_blobs = bottom_blobs;
- int ret = layer->forward_inplace(bottom_top_blobs, opt);
- if (ret != 0)
- return ret;
-
- // store top blobs
- for (size_t i = 0; i < layer->tops.size(); i++)
- {
- int top_blob_index = layer->tops[i];
-
- blob_mats[top_blob_index] = bottom_top_blobs[i];
- }
- }
- else
- {
- std::vector<Mat> top_blobs(layer->tops.size());
- int ret = layer->forward(bottom_blobs, top_blobs, opt);
- if (ret != 0)
- return ret;
-
- // store top blobs
- for (size_t i = 0; i < layer->tops.size(); i++)
- {
- int top_blob_index = layer->tops[i];
-
- blob_mats[top_blob_index] = top_blobs[i];
- }
- }
-
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats[bottom_blob_index].release();
- }
- }
- }
-
- return 0;
- }
-
- #if NCNN_VULKAN
- int NetPrivate::do_forward_layer(const Layer* layer, std::vector<VkMat>& blob_mats_gpu, VkCompute& cmd, const Option& opt) const
- {
- if (layer->one_blob_only)
- {
- // load bottom blob
- int bottom_blob_index = layer->bottoms[0];
- int top_blob_index = layer->tops[0];
-
- VkMat& bottom_blob_ref = blob_mats_gpu[bottom_blob_index];
- VkMat bottom_blob;
-
- if (opt.lightmode)
- {
- // deep copy for inplace forward if data is shared
- if (layer->support_inplace && *bottom_blob_ref.refcount != 1)
- {
- cmd.record_clone(bottom_blob_ref, bottom_blob, opt);
- // NCNN_LOGE("clone %p[+%lu] %p[+%lu]", bottom_blob_ref.buffer(), bottom_blob_ref.buffer_offset(), bottom_blob.buffer(), bottom_blob.buffer_offset());
- }
- }
- if (bottom_blob.dims == 0)
- {
- bottom_blob = bottom_blob_ref;
- }
-
- // forward
- if (opt.lightmode && layer->support_inplace)
- {
- VkMat& bottom_top_blob = bottom_blob;
- int ret = layer->forward_inplace(bottom_top_blob, cmd, opt);
- if (ret != 0)
- return ret;
-
- // store top blob
- blob_mats_gpu[top_blob_index] = bottom_top_blob;
- }
- else
- {
- VkMat top_blob;
- int ret = layer->forward(bottom_blob, top_blob, cmd, opt);
- if (ret != 0)
- return ret;
-
- // store top blob
- blob_mats_gpu[top_blob_index] = top_blob;
- }
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu[bottom_blob_index].release();
- }
- }
- else
- {
- // load bottom blobs
- std::vector<VkMat> bottom_blobs(layer->bottoms.size());
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- VkMat& bottom_blob_ref = blob_mats_gpu[bottom_blob_index];
- bottom_blobs[i].release();
-
- if (opt.lightmode)
- {
- // deep copy for inplace forward if data is shared
- if (layer->support_inplace && *bottom_blob_ref.refcount != 1)
- {
- cmd.record_clone(bottom_blob_ref, bottom_blobs[i], opt);
- // NCNN_LOGE("clone %p[+%lu] %p[+%lu]", bottom_blob_ref.buffer(), bottom_blob_ref.buffer_offset(), bottom_blobs[i].buffer(), bottom_blobs[i].buffer_offset());
- }
- }
- if (bottom_blobs[i].dims == 0)
- {
- bottom_blobs[i] = bottom_blob_ref;
- }
- }
-
- // forward
- if (opt.lightmode && layer->support_inplace)
- {
- std::vector<VkMat>& bottom_top_blobs = bottom_blobs;
- int ret = layer->forward_inplace(bottom_top_blobs, cmd, opt);
- if (ret != 0)
- return ret;
-
- // store top blobs
- for (size_t i = 0; i < layer->tops.size(); i++)
- {
- int top_blob_index = layer->tops[i];
-
- blob_mats_gpu[top_blob_index] = bottom_top_blobs[i];
- }
- }
- else
- {
- std::vector<VkMat> top_blobs(layer->tops.size());
- int ret = layer->forward(bottom_blobs, top_blobs, cmd, opt);
- if (ret != 0)
- return ret;
-
- // store top blobs
- for (size_t i = 0; i < layer->tops.size(); i++)
- {
- int top_blob_index = layer->tops[i];
-
- blob_mats_gpu[top_blob_index] = top_blobs[i];
- }
- }
-
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu[bottom_blob_index].release();
- }
- }
- }
-
- return 0;
- }
-
- int NetPrivate::do_forward_layer(const Layer* layer, std::vector<VkImageMat>& blob_mats_gpu_image, VkCompute& cmd, const Option& opt) const
- {
- if (layer->one_blob_only)
- {
- // load bottom blob
- int bottom_blob_index = layer->bottoms[0];
- int top_blob_index = layer->tops[0];
-
- VkImageMat& bottom_blob_ref = blob_mats_gpu_image[bottom_blob_index];
- VkImageMat bottom_blob;
-
- if (opt.lightmode)
- {
- // deep copy for inplace forward if data is shared
- if (layer->support_inplace && *bottom_blob_ref.refcount != 1)
- {
- cmd.record_clone(bottom_blob_ref, bottom_blob, opt);
- // NCNN_LOGE("clone %p[+%lu] %p[+%lu]", bottom_blob_ref.buffer(), bottom_blob_ref.buffer_offset(), bottom_blob.buffer(), bottom_blob.buffer_offset());
- }
- }
- if (bottom_blob.dims == 0)
- {
- bottom_blob = bottom_blob_ref;
- }
-
- // forward
- if (opt.lightmode && layer->support_inplace)
- {
- VkImageMat& bottom_top_blob = bottom_blob;
- int ret = layer->forward_inplace(bottom_top_blob, cmd, opt);
- if (ret != 0)
- return ret;
-
- // store top blob
- blob_mats_gpu_image[top_blob_index] = bottom_top_blob;
- }
- else
- {
- VkImageMat top_blob;
- int ret = layer->forward(bottom_blob, top_blob, cmd, opt);
- if (ret != 0)
- return ret;
-
- // store top blob
- blob_mats_gpu_image[top_blob_index] = top_blob;
- }
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu_image[bottom_blob_index].release();
- }
- }
- else
- {
- // load bottom blobs
- std::vector<VkImageMat> bottom_blobs(layer->bottoms.size());
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- VkImageMat& bottom_blob_ref = blob_mats_gpu_image[bottom_blob_index];
-
- if (opt.lightmode)
- {
- // deep copy for inplace forward if data is shared
- if (layer->support_inplace && *bottom_blob_ref.refcount != 1)
- {
- cmd.record_clone(bottom_blob_ref, bottom_blobs[i], opt);
- // NCNN_LOGE("clone %p[+%lu] %p[+%lu]", bottom_blob_ref.buffer(), bottom_blob_ref.buffer_offset(), bottom_blobs[i].buffer(), bottom_blobs[i].buffer_offset());
- }
- }
- if (bottom_blobs[i].dims == 0)
- {
- bottom_blobs[i] = bottom_blob_ref;
- }
- }
-
- // forward
- if (opt.lightmode && layer->support_inplace)
- {
- std::vector<VkImageMat>& bottom_top_blobs = bottom_blobs;
- int ret = layer->forward_inplace(bottom_top_blobs, cmd, opt);
- if (ret != 0)
- return ret;
-
- // store top blobs
- for (size_t i = 0; i < layer->tops.size(); i++)
- {
- int top_blob_index = layer->tops[i];
-
- blob_mats_gpu_image[top_blob_index] = bottom_top_blobs[i];
- }
- }
- else
- {
- std::vector<VkImageMat> top_blobs(layer->tops.size());
- int ret = layer->forward(bottom_blobs, top_blobs, cmd, opt);
- if (ret != 0)
- return ret;
-
- // store top blobs
- for (size_t i = 0; i < layer->tops.size(); i++)
- {
- int top_blob_index = layer->tops[i];
-
- blob_mats_gpu_image[top_blob_index] = top_blobs[i];
- }
- }
-
- for (size_t i = 0; i < layer->bottoms.size(); i++)
- {
- int bottom_blob_index = layer->bottoms[i];
-
- if (opt.lightmode)
- {
- // delete after taken in light mode
- blob_mats_gpu_image[bottom_blob_index].release();
- }
- }
- }
-
- return 0;
- }
- #endif // NCNN_VULKAN
-
- void NetPrivate::update_input_output_indexes()
- {
- input_blob_indexes.clear();
- output_blob_indexes.clear();
-
- for (size_t i = 0; i < layers.size(); i++)
- {
- if (layers[i]->typeindex == LayerType::Input)
- {
- int blob_index = layers[i]->tops[0];
- input_blob_indexes.push_back(blob_index);
- }
- }
-
- for (size_t i = 0; i < blobs.size(); i++)
- {
- if (blobs[i].producer != -1 && blobs[i].consumer == -1)
- {
- output_blob_indexes.push_back(i);
- }
- }
- }
-
- #if NCNN_STRING
- void NetPrivate::update_input_output_names()
- {
- input_blob_names.clear();
- output_blob_names.clear();
-
- for (size_t i = 0; i < input_blob_indexes.size(); i++)
- {
- int blob_index = input_blob_indexes[i];
- input_blob_names.push_back(blobs[blob_index].name.c_str());
- }
-
- for (size_t i = 0; i < output_blob_indexes.size(); i++)
- {
- int blob_index = output_blob_indexes[i];
- output_blob_names.push_back(blobs[blob_index].name.c_str());
- }
- }
- #endif // NCNN_STRING
-
- Net::Net()
- : d(new NetPrivate(opt))
- {
- }
-
- Net::~Net()
- {
- clear();
-
- delete d;
- }
-
- Net::Net(const Net&)
- : d(0)
- {
- }
-
- Net& Net::operator=(const Net&)
- {
- return *this;
- }
-
- #if NCNN_STRING
- int Net::register_custom_layer(const char* type, layer_creator_func creator, layer_destroyer_func destroyer, void* userdata)
- {
- int typeindex = layer_to_index(type);
- if (typeindex != -1)
- {
- NCNN_LOGE("can not register build-in layer type %s", type);
- return -1;
- }
-
- int custom_index = custom_layer_to_index(type);
- if (custom_index == -1)
- {
- struct custom_layer_registry_entry entry = {type, creator, destroyer, userdata};
- d->custom_layer_registry.push_back(entry);
- }
- else
- {
- NCNN_LOGE("overwrite existing custom layer type %s", type);
- d->custom_layer_registry[custom_index].name = type;
- d->custom_layer_registry[custom_index].creator = creator;
- d->custom_layer_registry[custom_index].destroyer = destroyer;
- d->custom_layer_registry[custom_index].userdata = userdata;
- }
-
- return 0;
- }
- #endif // NCNN_STRING
-
- int Net::register_custom_layer(int index, layer_creator_func creator, layer_destroyer_func destroyer, void* userdata)
- {
- int custom_index = index & ~LayerType::CustomBit;
- if (index == custom_index)
- {
- NCNN_LOGE("can not register build-in layer index %d", custom_index);
- return -1;
- }
-
- if ((int)d->custom_layer_registry.size() <= custom_index)
- {
- #if NCNN_STRING
- struct custom_layer_registry_entry dummy = {"", 0, 0, 0};
- #else
- struct custom_layer_registry_entry dummy = {0, 0, 0};
- #endif // NCNN_STRING
- d->custom_layer_registry.resize(custom_index + 1, dummy);
- }
-
- if (d->custom_layer_registry[custom_index].creator)
- {
- NCNN_LOGE("overwrite existing custom layer index %d", custom_index);
- }
-
- d->custom_layer_registry[custom_index].creator = creator;
- d->custom_layer_registry[custom_index].destroyer = destroyer;
- d->custom_layer_registry[custom_index].userdata = userdata;
- return 0;
- }
-
- #if NCNN_STRING
- int Net::load_param(const DataReader& dr)
- {
- #define SCAN_VALUE(fmt, v) \
- if (dr.scan(fmt, &v) != 1) \
- { \
- NCNN_LOGE("parse " #v " failed"); \
- return -1; \
- }
-
- int magic = 0;
- SCAN_VALUE("%d", magic)
- if (magic != 7767517)
- {
- NCNN_LOGE("param is too old, please regenerate");
- return -1;
- }
-
- // parse
- int layer_count = 0;
- int blob_count = 0;
- SCAN_VALUE("%d", layer_count)
- SCAN_VALUE("%d", blob_count)
- if (layer_count <= 0 || blob_count <= 0)
- {
- NCNN_LOGE("invalid layer_count or blob_count");
- return -1;
- }
-
- d->layers.resize((size_t)layer_count);
- d->blobs.resize((size_t)blob_count);
-
- #if NCNN_VULKAN
- // TODO enable gpu when bf16 conversion implemented
- if (opt.use_bf16_storage)
- opt.use_vulkan_compute = false;
-
- if (opt.use_vulkan_compute)
- {
- if (!d->vkdev) d->vkdev = get_gpu_device();
- if (!d->vkdev) opt.use_vulkan_compute = false; // no vulkan device, fallback to cpu
- }
- if (opt.use_vulkan_compute)
- {
- // sanitize use options
- if (!d->vkdev->info.support_fp16_packed()) opt.use_fp16_packed = false;
- if (!d->vkdev->info.support_fp16_storage()) opt.use_fp16_storage = false;
- if (!d->vkdev->info.support_fp16_arithmetic()) opt.use_fp16_arithmetic = false;
- if (!d->vkdev->info.support_int8_storage()) opt.use_int8_storage = false;
- if (!d->vkdev->info.support_int8_arithmetic()) opt.use_int8_arithmetic = false;
-
- if (d->vkdev->info.bug_buffer_image_load_zero()) opt.use_image_storage = false;
-
- // fp16a makes no sense when fp16 storage disabled
- if (!opt.use_fp16_packed && !opt.use_fp16_storage) opt.use_fp16_arithmetic = false;
- }
- else
- {
- // fp16a makes no sense when fp16 storage disabled
- if (!opt.use_fp16_storage) opt.use_fp16_arithmetic = false;
- }
- #endif // NCNN_VULKAN
-
- ParamDict pd;
-
- int blob_index = 0;
- for (int i = 0; i < layer_count; i++)
- {
- char layer_type[256];
- char layer_name[256];
- int bottom_count = 0;
- int top_count = 0;
- SCAN_VALUE("%255s", layer_type)
- SCAN_VALUE("%255s", layer_name)
- SCAN_VALUE("%d", bottom_count)
- SCAN_VALUE("%d", top_count)
-
- Layer* layer = create_layer(layer_type);
- if (!layer)
- {
- layer = create_custom_layer(layer_type);
- }
- if (!layer)
- {
- NCNN_LOGE("layer %s not exists or registered", layer_type);
- clear();
- return -1;
- }
-
- #if NCNN_VULKAN
- if (opt.use_vulkan_compute)
- layer->vkdev = d->vkdev;
- #endif // NCNN_VULKAN
-
- layer->type = std::string(layer_type);
- layer->name = std::string(layer_name);
- // NCNN_LOGE("new layer %d %s", i, layer_name);
-
- layer->bottoms.resize(bottom_count);
-
- for (int j = 0; j < bottom_count; j++)
- {
- char bottom_name[256];
- SCAN_VALUE("%255s", bottom_name)
-
- int bottom_blob_index = find_blob_index_by_name(bottom_name);
- if (bottom_blob_index == -1)
- {
- Blob& blob = d->blobs[blob_index];
-
- bottom_blob_index = blob_index;
-
- blob.name = std::string(bottom_name);
- // NCNN_LOGE("new blob %s", bottom_name);
-
- blob_index++;
- }
-
- Blob& blob = d->blobs[bottom_blob_index];
-
- blob.consumer = i;
-
- layer->bottoms[j] = bottom_blob_index;
- }
-
- layer->tops.resize(top_count);
- for (int j = 0; j < top_count; j++)
- {
- Blob& blob = d->blobs[blob_index];
-
- char blob_name[256];
- SCAN_VALUE("%255s", blob_name)
-
- blob.name = std::string(blob_name);
- // NCNN_LOGE("new blob %s", blob_name);
-
- blob.producer = i;
-
- layer->tops[j] = blob_index;
-
- blob_index++;
- }
-
- // layer specific params
- int pdlr = pd.load_param(dr);
- if (pdlr != 0)
- {
- NCNN_LOGE("ParamDict load_param %d %s failed", i, layer->name.c_str());
- continue;
- }
-
- if (layer->support_int8_storage)
- {
- // no int8 gpu support yet
- opt.use_vulkan_compute = false;
- }
-
- // pull out top shape hints
- Mat shape_hints = pd.get(30, Mat());
- if (!shape_hints.empty())
- {
- const int* psh = shape_hints;
- for (int j = 0; j < top_count; j++)
- {
- Blob& blob = d->blobs[layer->tops[j]];
-
- int dims = psh[0];
- if (dims == 1)
- {
- blob.shape = Mat(psh[1], (void*)0, 4u, 1);
- }
- if (dims == 2)
- {
- blob.shape = Mat(psh[1], psh[2], (void*)0, 4u, 1);
- }
- if (dims == 3)
- {
- blob.shape = Mat(psh[1], psh[2], psh[3], (void*)0, 4u, 1);
- }
-
- psh += 4;
- }
- }
-
- // set bottom and top shape hints
- layer->bottom_shapes.resize(bottom_count);
- for (int j = 0; j < bottom_count; j++)
- {
- layer->bottom_shapes[j] = d->blobs[layer->bottoms[j]].shape;
- }
-
- layer->top_shapes.resize(top_count);
- for (int j = 0; j < top_count; j++)
- {
- layer->top_shapes[j] = d->blobs[layer->tops[j]].shape;
- }
-
- int lr = layer->load_param(pd);
- if (lr != 0)
- {
- NCNN_LOGE("layer load_param %d %s failed", i, layer->name.c_str());
- continue;
- }
-
- d->layers[i] = layer;
- }
-
- d->update_input_output_indexes();
- d->update_input_output_names();
-
- #undef SCAN_VALUE
- return 0;
- }
- #endif // NCNN_STRING
-
- int Net::load_param_bin(const DataReader& dr)
- {
- #define READ_VALUE(buf) \
- if (dr.read(&buf, sizeof(buf)) != sizeof(buf)) \
- { \
- NCNN_LOGE("read " #buf " failed"); \
- return -1; \
- }
-
- int magic = 0;
- READ_VALUE(magic)
- if (magic != 7767517)
- {
- NCNN_LOGE("param is too old, please regenerate");
- return -1;
- }
-
- int layer_count = 0;
- int blob_count = 0;
- READ_VALUE(layer_count)
- READ_VALUE(blob_count)
- if (layer_count <= 0 || blob_count <= 0)
- {
- NCNN_LOGE("invalid layer_count or blob_count");
- return -1;
- }
-
- d->layers.resize(layer_count);
- d->blobs.resize(blob_count);
-
- #if NCNN_VULKAN
- // TODO enable gpu when bf16 conversion implemented
- if (opt.use_bf16_storage)
- opt.use_vulkan_compute = false;
-
- if (opt.use_vulkan_compute)
- {
- if (!d->vkdev) d->vkdev = get_gpu_device();
- if (!d->vkdev) opt.use_vulkan_compute = false; // no vulkan device, fallback to cpu
- }
- if (opt.use_vulkan_compute)
- {
- // sanitize use options
- if (!d->vkdev->info.support_fp16_packed()) opt.use_fp16_packed = false;
- if (!d->vkdev->info.support_fp16_storage()) opt.use_fp16_storage = false;
- if (!d->vkdev->info.support_fp16_arithmetic()) opt.use_fp16_arithmetic = false;
- if (!d->vkdev->info.support_int8_storage()) opt.use_int8_storage = false;
- if (!d->vkdev->info.support_int8_arithmetic()) opt.use_int8_arithmetic = false;
-
- if (d->vkdev->info.bug_buffer_image_load_zero()) opt.use_image_storage = false;
-
- // fp16a makes no sense when fp16 storage disabled
- if (!opt.use_fp16_packed && !opt.use_fp16_storage) opt.use_fp16_arithmetic = false;
- }
- else
- {
- // fp16a makes no sense when fp16 storage disabled
- if (!opt.use_fp16_storage) opt.use_fp16_arithmetic = false;
- }
- #endif // NCNN_VULKAN
-
- ParamDict pd;
-
- for (int i = 0; i < layer_count; i++)
- {
- int typeindex;
- int bottom_count;
- int top_count;
- READ_VALUE(typeindex)
- READ_VALUE(bottom_count)
- READ_VALUE(top_count)
-
- Layer* layer = create_layer(typeindex);
- if (!layer)
- {
- int custom_index = typeindex & ~LayerType::CustomBit;
- layer = create_custom_layer(custom_index);
- }
- if (!layer)
- {
- NCNN_LOGE("layer %d not exists or registered", typeindex);
- clear();
- return -1;
- }
-
- #if NCNN_VULKAN
- if (opt.use_vulkan_compute)
- layer->vkdev = d->vkdev;
- #endif // NCNN_VULKAN
-
- // layer->type = std::string(layer_type);
- // layer->name = std::string(layer_name);
- // NCNN_LOGE("new layer %d", typeindex);
-
- layer->bottoms.resize(bottom_count);
- for (int j = 0; j < bottom_count; j++)
- {
- int bottom_blob_index;
- READ_VALUE(bottom_blob_index)
-
- Blob& blob = d->blobs[bottom_blob_index];
-
- blob.consumer = i;
-
- layer->bottoms[j] = bottom_blob_index;
- }
-
- layer->tops.resize(top_count);
- for (int j = 0; j < top_count; j++)
- {
- int top_blob_index;
- READ_VALUE(top_blob_index)
-
- Blob& blob = d->blobs[top_blob_index];
-
- // blob.name = std::string(blob_name);
- // NCNN_LOGE("new blob %s", blob_name);
-
- blob.producer = i;
-
- layer->tops[j] = top_blob_index;
- }
-
- // layer specific params
- int pdlr = pd.load_param_bin(dr);
- if (pdlr != 0)
- {
- #if NCNN_STRING
- NCNN_LOGE("ParamDict load_param %d %s failed", i, layer->name.c_str());
- #else
- NCNN_LOGE("ParamDict load_param %d failed", i);
- #endif
- continue;
- }
-
- if (layer->support_int8_storage)
- {
- // no int8 gpu support yet
- opt.use_vulkan_compute = false;
- }
-
- // pull out top blob shape hints
- Mat shape_hints = pd.get(30, Mat());
- if (!shape_hints.empty())
- {
- const int* psh = shape_hints;
- for (int j = 0; j < top_count; j++)
- {
- Blob& blob = d->blobs[layer->tops[j]];
-
- int dims = psh[0];
- if (dims == 1)
- {
- blob.shape = Mat(psh[1], (void*)0, 4u, 1);
- }
- if (dims == 2)
- {
- blob.shape = Mat(psh[1], psh[2], (void*)0, 4u, 1);
- }
- if (dims == 3)
- {
- blob.shape = Mat(psh[1], psh[2], psh[3], (void*)0, 4u, 1);
- }
-
- psh += 4;
- }
- }
-
- // set bottom and top shape hints
- layer->bottom_shapes.resize(bottom_count);
- for (int j = 0; j < bottom_count; j++)
- {
- layer->bottom_shapes[j] = d->blobs[layer->bottoms[j]].shape;
- }
-
- layer->top_shapes.resize(top_count);
- for (int j = 0; j < top_count; j++)
- {
- layer->top_shapes[j] = d->blobs[layer->tops[j]].shape;
- }
-
- int lr = layer->load_param(pd);
- if (lr != 0)
- {
- #if NCNN_STRING
- NCNN_LOGE("layer load_param %d %s failed", i, layer->name.c_str());
- #else
- NCNN_LOGE("layer load_param %d failed", i);
- #endif
- continue;
- }
-
- d->layers[i] = layer;
- }
-
- d->update_input_output_indexes();
-
- #undef READ_VALUE
- return 0;
- }
-
- int Net::load_model(const DataReader& dr)
- {
- if (d->layers.empty())
- {
- NCNN_LOGE("network graph not ready");
- return -1;
- }
-
- int layer_count = (int)d->layers.size();
-
- // load file
- int ret = 0;
-
- ModelBinFromDataReader mb(dr);
- for (int i = 0; i < layer_count; i++)
- {
- Layer* layer = d->layers[i];
-
- //Here we found inconsistent content in the parameter file.
- if (!layer)
- {
- NCNN_LOGE("load_model error at layer %d, parameter file has inconsistent content.", i);
- ret = -1;
- break;
- }
-
- int lret = layer->load_model(mb);
- if (lret != 0)
- {
- #if NCNN_STRING
- NCNN_LOGE("layer load_model %d %s failed", i, layer->name.c_str());
- #else
- NCNN_LOGE("layer load_model %d failed", i);
- #endif
- ret = -1;
- break;
- }
-
- if (layer->support_int8_storage)
- {
- // no int8 gpu support yet
- opt.use_vulkan_compute = false;
- }
- }
-
- #if NCNN_VULKAN
- if (opt.use_vulkan_compute)
- {
- if (!opt.pipeline_cache)
- {
- if (!d->pipeline_cache)
- d->pipeline_cache = new PipelineCache(d->vkdev);
- opt.pipeline_cache = d->pipeline_cache;
- }
- }
- #endif // NCNN_VULKAN
-
- for (int i = 0; i < layer_count; i++)
- {
- Layer* layer = d->layers[i];
-
- Option opt1 = opt;
- #if NCNN_VULKAN
- if (opt.use_vulkan_compute)
- {
- if (!layer->support_image_storage) opt1.use_image_storage = false;
- }
- #endif // NCNN_VULKAN
-
- int cret = layer->create_pipeline(opt1);
- if (cret != 0)
- {
- #if NCNN_STRING
- NCNN_LOGE("layer create_pipeline %d %s failed", i, layer->name.c_str());
- #else
- NCNN_LOGE("layer create_pipeline %d failed", i);
- #endif
- ret = -1;
- break;
- }
- }
-
- if (opt.use_local_pool_allocator)
- {
- if (opt.blob_allocator == 0)
- {
- if (!d->local_blob_allocator)
- {
- d->local_blob_allocator = new PoolAllocator;
- d->local_blob_allocator->set_size_compare_ratio(0.f);
- }
- }
- if (opt.workspace_allocator == 0)
- {
- if (!d->local_workspace_allocator)
- {
- d->local_workspace_allocator = new PoolAllocator;
- d->local_workspace_allocator->set_size_compare_ratio(0.5f);
- }
- }
- }
-
- #if NCNN_VULKAN
- if (opt.use_vulkan_compute)
- {
- d->upload_model();
- }
- #endif // NCNN_VULKAN
-
- return ret;
- }
-
- #if NCNN_STDIO
- #if NCNN_STRING
- int Net::load_param(FILE* fp)
- {
- DataReaderFromStdio dr(fp);
- return load_param(dr);
- }
-
- int Net::load_param_mem(const char* _mem)
- {
- const unsigned char* mem = (const unsigned char*)_mem;
- DataReaderFromMemory dr(mem);
- return load_param(dr);
- }
-
- int Net::load_param(const char* protopath)
- {
- FILE* fp = fopen(protopath, "rb");
- if (!fp)
- {
- NCNN_LOGE("fopen %s failed", protopath);
- return -1;
- }
-
- int ret = load_param(fp);
- fclose(fp);
- return ret;
- }
- #endif // NCNN_STRING
-
- int Net::load_param_bin(FILE* fp)
- {
- DataReaderFromStdio dr(fp);
- return load_param_bin(dr);
- }
-
- int Net::load_param_bin(const char* protopath)
- {
- FILE* fp = fopen(protopath, "rb");
- if (!fp)
- {
- NCNN_LOGE("fopen %s failed", protopath);
- return -1;
- }
-
- int ret = load_param_bin(fp);
- fclose(fp);
- return ret;
- }
-
- int Net::load_model(FILE* fp)
- {
- DataReaderFromStdio dr(fp);
- return load_model(dr);
- }
-
- int Net::load_model(const char* modelpath)
- {
- FILE* fp = fopen(modelpath, "rb");
- if (!fp)
- {
- NCNN_LOGE("fopen %s failed", modelpath);
- return -1;
- }
-
- int ret = load_model(fp);
- fclose(fp);
- return ret;
- }
- #endif // NCNN_STDIO
-
- int Net::load_param(const unsigned char* _mem)
- {
- const unsigned char* mem = _mem;
- DataReaderFromMemory dr(mem);
- load_param_bin(dr);
- return static_cast<int>(mem - _mem);
- }
-
- int Net::load_model(const unsigned char* _mem)
- {
- const unsigned char* mem = _mem;
- DataReaderFromMemory dr(mem);
- load_model(dr);
- return static_cast<int>(mem - _mem);
- }
-
- #if NCNN_PLATFORM_API
- #if __ANDROID_API__ >= 9
- #if NCNN_STRING
- int Net::load_param(AAsset* asset)
- {
- DataReaderFromAndroidAsset dr(asset);
- return load_param(dr);
- }
-
- int Net::load_param(AAssetManager* mgr, const char* assetpath)
- {
- AAsset* asset = AAssetManager_open(mgr, assetpath, AASSET_MODE_BUFFER);
- if (!asset)
- {
- NCNN_LOGE("AAssetManager_open %s failed", assetpath);
- return -1;
- }
-
- int ret = load_param(asset);
- AAsset_close(asset);
- return ret;
- }
- #endif // NCNN_STRING
-
- int Net::load_param_bin(AAsset* asset)
- {
- DataReaderFromAndroidAsset dr(asset);
- return load_param_bin(dr);
- }
-
- int Net::load_param_bin(AAssetManager* mgr, const char* assetpath)
- {
- AAsset* asset = AAssetManager_open(mgr, assetpath, AASSET_MODE_BUFFER);
- if (!asset)
- {
- NCNN_LOGE("AAssetManager_open %s failed", assetpath);
- return -1;
- }
-
- int ret = load_param_bin(asset);
- AAsset_close(asset);
- return ret;
- }
-
- int Net::load_model(AAsset* asset)
- {
- DataReaderFromAndroidAsset dr(asset);
- return load_model(dr);
- }
-
- int Net::load_model(AAssetManager* mgr, const char* assetpath)
- {
- AAsset* asset = AAssetManager_open(mgr, assetpath, AASSET_MODE_STREAMING);
- if (!asset)
- {
- NCNN_LOGE("AAssetManager_open %s failed", assetpath);
- return -1;
- }
-
- int ret = load_model(asset);
- AAsset_close(asset);
- return ret;
- }
- #endif // __ANDROID_API__ >= 9
- #endif // NCNN_PLATFORM_API
-
- void Net::clear()
- {
- d->blobs.clear();
- for (size_t i = 0; i < d->layers.size(); i++)
- {
- Layer* layer = d->layers[i];
-
- Option opt1 = opt;
- if (!layer->support_image_storage)
- {
- opt1.use_image_storage = false;
- }
-
- int dret = layer->destroy_pipeline(opt1);
- if (dret != 0)
- {
- NCNN_LOGE("layer destroy_pipeline failed");
- // ignore anyway
- }
-
- if (layer->typeindex & ncnn::LayerType::CustomBit)
- {
- int custom_index = layer->typeindex & ~ncnn::LayerType::CustomBit;
- if (d->custom_layer_registry[custom_index].destroyer)
- {
- d->custom_layer_registry[custom_index].destroyer(layer, d->custom_layer_registry[custom_index].userdata);
- }
- else
- {
- delete layer;
- }
- }
- else
- {
- delete layer;
- }
- }
- d->layers.clear();
-
- if (d->local_blob_allocator)
- {
- delete d->local_blob_allocator;
- d->local_blob_allocator = 0;
- }
- if (d->local_workspace_allocator)
- {
- delete d->local_workspace_allocator;
- d->local_workspace_allocator = 0;
- }
-
- #if NCNN_VULKAN
- if (d->weight_vkallocator)
- {
- delete d->weight_vkallocator;
- d->weight_vkallocator = 0;
- }
- if (d->weight_staging_vkallocator)
- {
- delete d->weight_staging_vkallocator;
- d->weight_staging_vkallocator = 0;
- }
- if (d->pipeline_cache)
- {
- delete d->pipeline_cache;
- d->pipeline_cache = 0;
- opt.pipeline_cache = 0;
- }
- #endif // NCNN_VULKAN
- }
-
- Extractor Net::create_extractor() const
- {
- return Extractor(this, d->blobs.size());
- }
-
- const std::vector<int>& Net::input_indexes() const
- {
- return d->input_blob_indexes;
- }
-
- const std::vector<int>& Net::output_indexes() const
- {
- return d->output_blob_indexes;
- }
-
- #if NCNN_STRING
- const std::vector<const char*>& Net::input_names() const
- {
- return d->input_blob_names;
- }
-
- const std::vector<const char*>& Net::output_names() const
- {
- return d->output_blob_names;
- }
- #endif
-
- const std::vector<Blob>& Net::blobs() const
- {
- return d->blobs;
- }
-
- const std::vector<Layer*>& Net::layers() const
- {
- return d->layers;
- }
-
- std::vector<Blob>& Net::mutable_blobs()
- {
- return d->blobs;
- }
-
- std::vector<Layer*>& Net::mutable_layers()
- {
- return d->layers;
- }
-
- #if NCNN_VULKAN
- void Net::set_vulkan_device(int device_index)
- {
- d->vkdev = get_gpu_device(device_index);
- }
-
- void Net::set_vulkan_device(const VulkanDevice* _vkdev)
- {
- d->vkdev = _vkdev;
- }
-
- const VulkanDevice* Net::vulkan_device() const
- {
- return d->vkdev;
- }
- #endif // NCNN_VULKAN
-
- #if NCNN_STRING
- int Net::find_blob_index_by_name(const char* name) const
- {
- for (size_t i = 0; i < d->blobs.size(); i++)
- {
- const Blob& blob = d->blobs[i];
- if (blob.name == name)
- {
- return static_cast<int>(i);
- }
- }
-
- NCNN_LOGE("find_blob_index_by_name %s failed", name);
- return -1;
- }
-
- int Net::find_layer_index_by_name(const char* name) const
- {
- for (size_t i = 0; i < d->layers.size(); i++)
- {
- const Layer* layer = d->layers[i];
- if (layer->name == name)
- {
- return static_cast<int>(i);
- }
- }
-
- NCNN_LOGE("find_layer_index_by_name %s failed", name);
- return -1;
- }
-
- int Net::custom_layer_to_index(const char* type)
- {
- const size_t custom_layer_registry_entry_count = d->custom_layer_registry.size();
- for (size_t i = 0; i < custom_layer_registry_entry_count; i++)
- {
- if (strcmp(type, d->custom_layer_registry[i].name) == 0)
- return static_cast<int>(i);
- }
-
- return -1;
- }
-
- Layer* Net::create_custom_layer(const char* type)
- {
- int index = custom_layer_to_index(type);
- if (index == -1)
- return 0;
-
- return create_custom_layer(index);
- }
- #endif // NCNN_STRING
-
- Layer* Net::create_custom_layer(int index)
- {
- const size_t custom_layer_registry_entry_count = d->custom_layer_registry.size();
- if (index < 0 || static_cast<unsigned int>(index) >= custom_layer_registry_entry_count)
- return 0;
-
- layer_creator_func layer_creator = d->custom_layer_registry[index].creator;
- if (!layer_creator)
- return 0;
-
- Layer* layer = layer_creator(d->custom_layer_registry[index].userdata);
- layer->typeindex = ncnn::LayerType::CustomBit | index;
- return layer;
- }
-
- class ExtractorPrivate
- {
- public:
- ExtractorPrivate(const Net* _net)
- : net(_net)
- {
- }
- const Net* net;
- std::vector<Mat> blob_mats;
- Option opt;
-
- #if NCNN_VULKAN
- VkAllocator* local_blob_vkallocator;
- VkAllocator* local_staging_vkallocator;
-
- std::vector<VkMat> blob_mats_gpu;
- std::vector<VkImageMat> blob_mats_gpu_image;
- #endif // NCNN_VULKAN
- };
-
- Extractor::Extractor(const Net* _net, size_t blob_count)
- : d(new ExtractorPrivate(_net))
- {
- d->blob_mats.resize(blob_count);
- d->opt = d->net->opt;
-
- #if NCNN_VULKAN
- if (d->net->opt.use_vulkan_compute)
- {
- d->local_blob_vkallocator = 0;
- d->local_staging_vkallocator = 0;
-
- d->blob_mats_gpu.resize(blob_count);
- d->blob_mats_gpu_image.resize(blob_count);
- }
- #endif // NCNN_VULKAN
- }
-
- Extractor::~Extractor()
- {
- clear();
-
- delete d;
- }
-
- Extractor::Extractor(const Extractor& rhs)
- : d(new ExtractorPrivate(0))
- {
- d->net = rhs.d->net;
- d->blob_mats = rhs.d->blob_mats;
- d->opt = rhs.d->opt;
-
- #if NCNN_VULKAN
- d->local_blob_vkallocator = 0;
- d->local_staging_vkallocator = 0;
-
- d->blob_mats_gpu = rhs.d->blob_mats_gpu;
- d->blob_mats_gpu_image = rhs.d->blob_mats_gpu_image;
- #endif // NCNN_VULKAN
- }
-
- Extractor& Extractor::operator=(const Extractor& rhs)
- {
- if (this == &rhs)
- return *this;
-
- d->net = rhs.d->net;
- d->blob_mats = rhs.d->blob_mats;
- d->opt = rhs.d->opt;
-
- #if NCNN_VULKAN
- d->local_blob_vkallocator = 0;
- d->local_staging_vkallocator = 0;
-
- d->blob_mats_gpu = rhs.d->blob_mats_gpu;
- d->blob_mats_gpu_image = rhs.d->blob_mats_gpu_image;
- #endif // NCNN_VULKAN
-
- return *this;
- }
-
- void Extractor::clear()
- {
- d->blob_mats.clear();
-
- #if NCNN_VULKAN
- if (d->opt.use_vulkan_compute)
- {
- d->blob_mats_gpu.clear();
- d->blob_mats_gpu_image.clear();
-
- if (d->local_blob_vkallocator)
- {
- d->net->vulkan_device()->reclaim_blob_allocator(d->local_blob_vkallocator);
- }
- if (d->local_staging_vkallocator)
- {
- d->net->vulkan_device()->reclaim_staging_allocator(d->local_staging_vkallocator);
- }
- }
- #endif // NCNN_VULKAN
- }
-
- void Extractor::set_light_mode(bool enable)
- {
- d->opt.lightmode = enable;
- }
-
- void Extractor::set_num_threads(int num_threads)
- {
- d->opt.num_threads = num_threads;
- }
-
- void Extractor::set_blob_allocator(Allocator* allocator)
- {
- d->opt.blob_allocator = allocator;
- }
-
- void Extractor::set_workspace_allocator(Allocator* allocator)
- {
- d->opt.workspace_allocator = allocator;
- }
-
- #if NCNN_VULKAN
- void Extractor::set_vulkan_compute(bool enable)
- {
- if (d->net->d->opt.use_vulkan_compute)
- {
- d->opt.use_vulkan_compute = enable;
- }
- else
- {
- NCNN_LOGE("set_vulkan_compute failed, network use_vulkan_compute disabled");
- }
- }
-
- void Extractor::set_blob_vkallocator(VkAllocator* allocator)
- {
- d->opt.blob_vkallocator = allocator;
- }
-
- void Extractor::set_workspace_vkallocator(VkAllocator* allocator)
- {
- d->opt.workspace_vkallocator = allocator;
- }
-
- void Extractor::set_staging_vkallocator(VkAllocator* allocator)
- {
- d->opt.staging_vkallocator = allocator;
- }
- #endif // NCNN_VULKAN
-
- #if NCNN_STRING
- int Extractor::input(const char* blob_name, const Mat& in)
- {
- int blob_index = d->net->find_blob_index_by_name(blob_name);
- if (blob_index == -1)
- {
- NCNN_LOGE("Try");
- const std::vector<const char*>& input_names = d->net->input_names();
- for (size_t i = 0; i < input_names.size(); i++)
- {
- NCNN_LOGE(" ex.input(\"%s\", in%d);", input_names[i], (int)i);
- }
-
- return -1;
- }
-
- return input(blob_index, in);
- }
-
- int Extractor::extract(const char* blob_name, Mat& feat, int type)
- {
- int blob_index = d->net->find_blob_index_by_name(blob_name);
- if (blob_index == -1)
- {
- NCNN_LOGE("Try");
- const std::vector<const char*>& output_names = d->net->output_names();
- for (size_t i = 0; i < output_names.size(); i++)
- {
- NCNN_LOGE(" ex.extract(\"%s\", out%d);", output_names[i], (int)i);
- }
-
- return -1;
- }
-
- return extract(blob_index, feat, type);
- }
- #endif // NCNN_STRING
-
- int Extractor::input(int blob_index, const Mat& in)
- {
- if (blob_index < 0 || blob_index >= (int)d->blob_mats.size())
- return -1;
-
- d->blob_mats[blob_index] = in;
-
- return 0;
- }
-
- int Extractor::extract(int blob_index, Mat& feat, int type)
- {
- if (blob_index < 0 || blob_index >= (int)d->blob_mats.size())
- return -1;
-
- int old_blocktime = get_kmp_blocktime();
- set_kmp_blocktime(d->opt.openmp_blocktime);
-
- int old_flush_denormals = get_flush_denormals();
- set_flush_denormals(d->opt.flush_denormals);
-
- int ret = 0;
-
- if (d->blob_mats[blob_index].dims == 0)
- {
- int layer_index = d->net->blobs()[blob_index].producer;
-
- // use local allocator
- if (d->opt.use_local_pool_allocator)
- {
- if (!d->opt.blob_allocator)
- {
- d->opt.blob_allocator = d->net->d->local_blob_allocator;
- }
- if (!d->opt.workspace_allocator)
- {
- d->opt.workspace_allocator = d->net->d->local_workspace_allocator;
- }
- }
-
- #if NCNN_VULKAN
- if (d->opt.use_vulkan_compute)
- {
- // use local allocator
- if (!d->opt.blob_vkallocator)
- {
- d->local_blob_vkallocator = d->net->vulkan_device()->acquire_blob_allocator();
- d->opt.blob_vkallocator = d->local_blob_vkallocator;
- }
- if (!d->opt.workspace_vkallocator)
- {
- d->opt.workspace_vkallocator = d->opt.blob_vkallocator;
- }
- if (!d->opt.staging_vkallocator)
- {
- d->local_staging_vkallocator = d->net->vulkan_device()->acquire_staging_allocator();
- d->opt.staging_vkallocator = d->local_staging_vkallocator;
- }
-
- ncnn::VkCompute cmd(d->net->vulkan_device());
- #if NCNN_BENCHMARK
- cmd.create_query_pool(d->net->layers().size() * 2);
- #endif // NCNN_BENCHMARK
-
- // TODO vkimagemat for adreno
- if (d->opt.use_image_storage)
- {
- VkImageMat feat_gpu;
- ret = extract(blob_index, feat_gpu, cmd);
-
- if (d->blob_mats[blob_index].dims == 0 && feat_gpu.dims != 0)
- {
- cmd.record_download(feat_gpu, d->blob_mats[blob_index], d->opt);
-
- cmd.submit_and_wait();
-
- #if NCNN_BENCHMARK
- std::vector<uint64_t> results(d->net->layers().size() * 2);
- cmd.get_query_pool_results(0, d->net->layers().size() * 2, results);
- for (size_t i = 0; i < d->net->layers().size(); i++)
- {
- uint64_t start = results[i * 2];
- uint64_t end = results[i * 2 + 1];
- if (start == 0 || end == 0)
- continue;
-
- double duration_us = (end - start) * d->net->vulkan_device()->info.timestamp_period() / 1000;
- NCNN_LOGE("%-24s %-30s %8.2lfus |", d->net->layers()[i]->type.c_str(), d->net->layers()[i]->name.c_str(), duration_us);
- }
- #endif // NCNN_BENCHMARK
- }
- }
- else
- {
- VkMat feat_gpu;
- ret = extract(blob_index, feat_gpu, cmd);
-
- if (d->blob_mats[blob_index].dims == 0 && feat_gpu.dims != 0)
- {
- cmd.record_download(feat_gpu, d->blob_mats[blob_index], d->opt);
-
- cmd.submit_and_wait();
-
- #if NCNN_BENCHMARK
- std::vector<uint64_t> results(d->net->layers().size() * 2);
- cmd.get_query_pool_results(0, d->net->layers().size() * 2, results);
- for (size_t i = 0; i < d->net->layers().size(); i++)
- {
- uint64_t start = results[i * 2];
- uint64_t end = results[i * 2 + 1];
- if (start == 0 || end == 0)
- continue;
-
- double duration_us = (end - start) * d->net->vulkan_device()->info.timestamp_period() / 1000;
- NCNN_LOGE("%-24s %-30s %8.2lfus |", d->net->layers()[i]->type.c_str(), d->net->layers()[i]->name.c_str(), duration_us);
- }
- #endif // NCNN_BENCHMARK
- }
- }
- }
- else
- {
- ret = d->net->d->forward_layer(layer_index, d->blob_mats, d->opt);
- }
- #else
- ret = d->net->d->forward_layer(layer_index, d->blob_mats, d->opt);
- #endif // NCNN_VULKAN
- }
-
- feat = d->blob_mats[blob_index];
-
- if (d->opt.use_packing_layout && (type == 0) && feat.elempack != 1)
- {
- Mat bottom_blob_unpacked;
- convert_packing(feat, bottom_blob_unpacked, 1, d->opt);
- feat = bottom_blob_unpacked;
- }
-
- // clang-format off
- // *INDENT-OFF*
- #if NCNN_ARM82
- if (d->opt.use_fp16_storage && cpu_support_arm_asimdhp() && (type == 0))
- {
- if (feat.elembits() == 16)
- {
- Mat feat_fp32;
- cast_float16_to_float32(feat, feat_fp32, d->opt);
- feat = feat_fp32;
- }
- }
- else
- #endif // NCNN_ARM82
- #if NCNN_BF16
- if (d->opt.use_bf16_storage && (type == 0))
- {
- if (feat.elembits() == 16)
- {
- Mat feat_fp32;
- cast_bfloat16_to_float32(feat, feat_fp32, d->opt);
- feat = feat_fp32;
- }
- }
- else
- #endif // NCNN_BF16
- if (feat.elembits() == 8 && (type == 0))
- {
- Mat feat_fp32;
- cast_int8_to_float32(feat, feat_fp32, d->opt);
- feat = feat_fp32;
- }
- // *INDENT-ON*
- // clang-format on
-
- if (d->opt.use_local_pool_allocator && feat.allocator == d->net->d->local_blob_allocator)
- {
- // detach the returned mat from local pool allocator
- // so we could destroy net instance much earlier
- feat = feat.clone();
- }
-
- set_kmp_blocktime(old_blocktime);
- set_flush_denormals(old_flush_denormals);
-
- return ret;
- }
-
- #if NCNN_VULKAN
- #if NCNN_STRING
- int Extractor::input(const char* blob_name, const VkMat& in)
- {
- int blob_index = d->net->find_blob_index_by_name(blob_name);
- if (blob_index == -1)
- {
- NCNN_LOGE("Try");
- const std::vector<const char*>& input_names = d->net->input_names();
- for (size_t i = 0; i < input_names.size(); i++)
- {
- NCNN_LOGE(" ex.input(\"%s\", in%d);", input_names[i], (int)i);
- }
-
- return -1;
- }
-
- return input(blob_index, in);
- }
-
- int Extractor::extract(const char* blob_name, VkMat& feat, VkCompute& cmd)
- {
- int blob_index = d->net->find_blob_index_by_name(blob_name);
- if (blob_index == -1)
- {
- NCNN_LOGE("Try");
- const std::vector<const char*>& output_names = d->net->output_names();
- for (size_t i = 0; i < output_names.size(); i++)
- {
- NCNN_LOGE(" ex.extract(\"%s\", out%d);", output_names[i], (int)i);
- }
-
- return -1;
- }
-
- return extract(blob_index, feat, cmd);
- }
-
- int Extractor::input(const char* blob_name, const VkImageMat& in)
- {
- int blob_index = d->net->find_blob_index_by_name(blob_name);
- if (blob_index == -1)
- {
- NCNN_LOGE("Try");
- const std::vector<const char*>& input_names = d->net->input_names();
- for (size_t i = 0; i < input_names.size(); i++)
- {
- NCNN_LOGE(" ex.input(\"%s\", in%d);", input_names[i], (int)i);
- }
-
- return -1;
- }
-
- return input(blob_index, in);
- }
-
- int Extractor::extract(const char* blob_name, VkImageMat& feat, VkCompute& cmd)
- {
- int blob_index = d->net->find_blob_index_by_name(blob_name);
- if (blob_index == -1)
- {
- NCNN_LOGE("Try");
- const std::vector<const char*>& output_names = d->net->output_names();
- for (size_t i = 0; i < output_names.size(); i++)
- {
- NCNN_LOGE(" ex.extract(\"%s\", out%d);", output_names[i], (int)i);
- }
-
- return -1;
- }
-
- return extract(blob_index, feat, cmd);
- }
- #endif // NCNN_STRING
-
- int Extractor::input(int blob_index, const VkMat& in)
- {
- if (blob_index < 0 || blob_index >= (int)d->blob_mats.size())
- return -1;
-
- d->blob_mats_gpu[blob_index] = in;
-
- return 0;
- }
-
- int Extractor::extract(int blob_index, VkMat& feat, VkCompute& cmd)
- {
- if (blob_index < 0 || blob_index >= (int)d->blob_mats.size())
- return -1;
-
- int old_blocktime = get_kmp_blocktime();
- set_kmp_blocktime(d->opt.openmp_blocktime);
-
- int old_flush_denormals = get_flush_denormals();
- set_flush_denormals(d->opt.flush_denormals);
-
- int ret = 0;
-
- if (d->blob_mats_gpu[blob_index].dims == 0)
- {
- if (d->blob_mats_gpu_image[blob_index].dims != 0)
- {
- // image to buffer
- cmd.record_image_to_buffer(d->blob_mats_gpu_image[blob_index], d->blob_mats_gpu[blob_index], d->opt);
- }
- else if (d->blob_mats[blob_index].dims != 0)
- {
- // host to buffer
- cmd.record_upload(d->blob_mats[blob_index], d->blob_mats_gpu[blob_index], d->opt);
- }
- else
- {
- int layer_index = d->net->blobs()[blob_index].producer;
- ret = d->net->d->forward_layer(layer_index, d->blob_mats, d->blob_mats_gpu, cmd, d->opt);
- }
- }
-
- feat = d->blob_mats_gpu[blob_index];
-
- set_kmp_blocktime(old_blocktime);
- set_flush_denormals(old_flush_denormals);
-
- return ret;
- }
-
- int Extractor::input(int blob_index, const VkImageMat& in)
- {
- if (blob_index < 0 || blob_index >= (int)d->blob_mats.size())
- return -1;
-
- d->blob_mats_gpu_image[blob_index] = in;
-
- return 0;
- }
-
- int Extractor::extract(int blob_index, VkImageMat& feat, VkCompute& cmd)
- {
- if (blob_index < 0 || blob_index >= (int)d->blob_mats.size())
- return -1;
-
- int old_blocktime = get_kmp_blocktime();
- set_kmp_blocktime(d->opt.openmp_blocktime);
-
- int old_flush_denormals = get_flush_denormals();
- set_flush_denormals(d->opt.flush_denormals);
-
- int ret = 0;
-
- if (d->blob_mats_gpu_image[blob_index].dims == 0)
- {
- if (d->blob_mats_gpu[blob_index].dims != 0)
- {
- // buffer to image
- cmd.record_buffer_to_image(d->blob_mats_gpu[blob_index], d->blob_mats_gpu_image[blob_index], d->opt);
- }
- else if (d->blob_mats[blob_index].dims != 0)
- {
- // host to image
- cmd.record_upload(d->blob_mats[blob_index], d->blob_mats_gpu_image[blob_index], d->opt);
- }
- else
- {
- int layer_index = d->net->blobs()[blob_index].producer;
- ret = d->net->d->forward_layer(layer_index, d->blob_mats, d->blob_mats_gpu, d->blob_mats_gpu_image, cmd, d->opt);
- }
- }
-
- feat = d->blob_mats_gpu_image[blob_index];
-
- if (feat.empty())
- {
- NCNN_LOGE("extract %d image allocation failed", blob_index);
- ret = -100;
- }
-
- set_kmp_blocktime(old_blocktime);
- set_flush_denormals(old_flush_denormals);
-
- return ret;
- }
- #endif // NCNN_VULKAN
-
- } // namespace ncnn
|