You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

benchncnn.cpp 13 kB

[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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595
  1. // Tencent is pleased to support the open source community by making ncnn available.
  2. //
  3. // Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
  4. //
  5. // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
  6. // in compliance with the License. You may obtain a copy of the License at
  7. //
  8. // https://opensource.org/licenses/BSD-3-Clause
  9. //
  10. // Unless required by applicable law or agreed to in writing, software distributed
  11. // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
  12. // CONDITIONS OF ANY KIND, either express or implied. See the License for the
  13. // specific language governing permissions and limitations under the License.
  14. #include <float.h>
  15. #include <stdio.h>
  16. #ifdef _WIN32
  17. #define NOMINMAX
  18. #include <algorithm>
  19. #include <windows.h> // Sleep()
  20. #else
  21. #include <unistd.h> // sleep()
  22. #endif
  23. #include "benchmark.h"
  24. #include "cpu.h"
  25. #include "net.h"
  26. #if NCNN_VULKAN
  27. #include "gpu.h"
  28. class GlobalGpuInstance
  29. {
  30. public:
  31. GlobalGpuInstance() { ncnn::create_gpu_instance(); }
  32. ~GlobalGpuInstance() { ncnn::destroy_gpu_instance(); }
  33. };
  34. // initialize vulkan runtime before main()
  35. GlobalGpuInstance g_global_gpu_instance;
  36. #endif // NCNN_VULKAN
  37. namespace ncnn {
  38. // always return empty weights
  39. class ModelBinFromEmpty : public ModelBin
  40. {
  41. public:
  42. virtual Mat load(int w, int /*type*/) const { return Mat(w); }
  43. };
  44. class BenchNet : public Net
  45. {
  46. public:
  47. int load_model()
  48. {
  49. // load file
  50. int ret = 0;
  51. ModelBinFromEmpty mb;
  52. for (size_t i=0; i<layers.size(); i++)
  53. {
  54. Layer* layer = layers[i];
  55. int lret = layer->load_model(mb);
  56. if (lret != 0)
  57. {
  58. fprintf(stderr, "layer load_model %d failed\n", (int)i);
  59. ret = -1;
  60. break;
  61. }
  62. }
  63. #if NCNN_VULKAN
  64. if (use_vulkan_compute)
  65. {
  66. upload_model();
  67. create_pipeline();
  68. }
  69. #endif // NCNN_VULKAN
  70. return ret;
  71. }
  72. };
  73. } // namespace ncnn
  74. static int g_loop_count = 4;
  75. static ncnn::UnlockedPoolAllocator g_blob_pool_allocator;
  76. static ncnn::PoolAllocator g_workspace_pool_allocator;
  77. #if NCNN_VULKAN
  78. static bool g_use_vulkan_compute = false;
  79. static ncnn::VulkanDevice* g_vkdev = 0;
  80. static ncnn::VkAllocator* g_blob_vkallocator = 0;
  81. static ncnn::VkAllocator* g_staging_vkallocator = 0;
  82. #endif // NCNN_VULKAN
  83. void benchmark(const char* comment, void (*init)(ncnn::Net&), void (*run)(const ncnn::Net&))
  84. {
  85. ncnn::BenchNet net;
  86. #if NCNN_VULKAN
  87. if (g_use_vulkan_compute)
  88. {
  89. net.use_vulkan_compute = g_use_vulkan_compute;
  90. net.set_vulkan_device(g_vkdev);
  91. }
  92. #endif // NCNN_VULKAN
  93. init(net);
  94. net.load_model();
  95. g_blob_pool_allocator.clear();
  96. g_workspace_pool_allocator.clear();
  97. #if NCNN_VULKAN
  98. if (g_use_vulkan_compute)
  99. {
  100. g_blob_vkallocator->clear();
  101. g_staging_vkallocator->clear();
  102. }
  103. #endif // NCNN_VULKAN
  104. // sleep 10 seconds for cooling down SOC :(
  105. #ifdef _WIN32
  106. Sleep(10 * 1000);
  107. #else
  108. sleep(10);
  109. #endif
  110. // warm up
  111. run(net);
  112. run(net);
  113. run(net);
  114. double time_min = DBL_MAX;
  115. double time_max = -DBL_MAX;
  116. double time_avg = 0;
  117. for (int i=0; i<g_loop_count; i++)
  118. {
  119. double start = ncnn::get_current_time();
  120. run(net);
  121. double end = ncnn::get_current_time();
  122. double time = end - start;
  123. time_min = std::min(time_min, time);
  124. time_max = std::max(time_max, time);
  125. time_avg += time;
  126. }
  127. time_avg /= g_loop_count;
  128. fprintf(stderr, "%20s min = %7.2f max = %7.2f avg = %7.2f\n", comment, time_min, time_max, time_avg);
  129. }
  130. void squeezenet_init(ncnn::Net& net)
  131. {
  132. net.load_param("squeezenet.param");
  133. }
  134. void squeezenet_int8_init(ncnn::Net& net)
  135. {
  136. net.load_param("squeezenet_int8.param");
  137. }
  138. void squeezenet_run(const ncnn::Net& net)
  139. {
  140. ncnn::Extractor ex = net.create_extractor();
  141. ncnn::Mat in(227, 227, 3);
  142. ex.input("data", in);
  143. ncnn::Mat out;
  144. ex.extract("prob", out);
  145. }
  146. void mobilenet_init(ncnn::Net& net)
  147. {
  148. net.load_param("mobilenet.param");
  149. }
  150. void mobilenet_int8_init(ncnn::Net& net)
  151. {
  152. net.load_param("mobilenet_int8.param");
  153. }
  154. void mobilenet_run(const ncnn::Net& net)
  155. {
  156. ncnn::Extractor ex = net.create_extractor();
  157. ncnn::Mat in(224, 224, 3);
  158. ex.input("data", in);
  159. ncnn::Mat out;
  160. ex.extract("prob", out);
  161. }
  162. void mobilenet_v2_init(ncnn::Net& net)
  163. {
  164. net.load_param("mobilenet_v2.param");
  165. }
  166. void mobilenet_v2_int8_init(ncnn::Net& net)
  167. {
  168. net.load_param("mobilenet_v2_int8.param");
  169. }
  170. void mobilenet_v2_run(const ncnn::Net& net)
  171. {
  172. ncnn::Extractor ex = net.create_extractor();
  173. ncnn::Mat in(224, 224, 3);
  174. ex.input("data", in);
  175. ncnn::Mat out;
  176. ex.extract("prob", out);
  177. }
  178. void shufflenet_init(ncnn::Net& net)
  179. {
  180. net.load_param("shufflenet.param");
  181. }
  182. void shufflenet_run(const ncnn::Net& net)
  183. {
  184. ncnn::Extractor ex = net.create_extractor();
  185. ncnn::Mat in(224, 224, 3);
  186. ex.input("data", in);
  187. ncnn::Mat out;
  188. ex.extract("fc1000", out);
  189. }
  190. void mnasnet_init(ncnn::Net& net)
  191. {
  192. net.load_param("mnasnet.param");
  193. }
  194. void mnasnet_run(const ncnn::Net& net)
  195. {
  196. ncnn::Extractor ex = net.create_extractor();
  197. ncnn::Mat in(224, 224, 3);
  198. ex.input("data", in);
  199. ncnn::Mat out;
  200. ex.extract("prob", out);
  201. }
  202. void proxylessnasnet_init(ncnn::Net& net)
  203. {
  204. net.load_param("proxylessnasnet.param");
  205. }
  206. void proxylessnasnet_run(const ncnn::Net& net)
  207. {
  208. ncnn::Extractor ex = net.create_extractor();
  209. ncnn::Mat in(224, 224, 3);
  210. ex.input("data", in);
  211. ncnn::Mat out;
  212. ex.extract("prob", out);
  213. }
  214. void googlenet_init(ncnn::Net& net)
  215. {
  216. net.load_param("googlenet.param");
  217. }
  218. void googlenet_int8_init(ncnn::Net& net)
  219. {
  220. net.load_param("googlenet_int8.param");
  221. }
  222. void googlenet_run(const ncnn::Net& net)
  223. {
  224. ncnn::Extractor ex = net.create_extractor();
  225. ncnn::Mat in(224, 224, 3);
  226. ex.input("data", in);
  227. ncnn::Mat out;
  228. ex.extract("prob", out);
  229. }
  230. void resnet18_init(ncnn::Net& net)
  231. {
  232. net.load_param("resnet18.param");
  233. }
  234. void resnet18_int8_init(ncnn::Net& net)
  235. {
  236. net.load_param("resnet18_int8.param");
  237. }
  238. void resnet18_run(const ncnn::Net& net)
  239. {
  240. ncnn::Extractor ex = net.create_extractor();
  241. ncnn::Mat in(224, 224, 3);
  242. ex.input("data", in);
  243. ncnn::Mat out;
  244. ex.extract("prob", out);
  245. }
  246. void alexnet_init(ncnn::Net& net)
  247. {
  248. net.load_param("alexnet.param");
  249. }
  250. void alexnet_run(const ncnn::Net& net)
  251. {
  252. ncnn::Extractor ex = net.create_extractor();
  253. ncnn::Mat in(227, 227, 3);
  254. ex.input("data", in);
  255. ncnn::Mat out;
  256. ex.extract("prob", out);
  257. }
  258. void vgg16_init(ncnn::Net& net)
  259. {
  260. net.load_param("vgg16.param");
  261. }
  262. void vgg16_int8_init(ncnn::Net& net)
  263. {
  264. net.load_param("vgg16_int8.param");
  265. }
  266. void vgg16_run(const ncnn::Net& net)
  267. {
  268. ncnn::Extractor ex = net.create_extractor();
  269. ncnn::Mat in(224, 224, 3);
  270. ex.input("data", in);
  271. ncnn::Mat out;
  272. ex.extract("prob", out);
  273. }
  274. void resnet50_init(ncnn::Net& net)
  275. {
  276. net.load_param("resnet50.param");
  277. }
  278. void resnet50_int8_init(ncnn::Net& net)
  279. {
  280. net.load_param("resnet50_int8.param");
  281. }
  282. void resnet50_run(const ncnn::Net& net)
  283. {
  284. ncnn::Extractor ex = net.create_extractor();
  285. ncnn::Mat in(224, 224, 3);
  286. ex.input("data", in);
  287. ncnn::Mat out;
  288. ex.extract("prob", out);
  289. }
  290. void squeezenet_ssd_init(ncnn::Net& net)
  291. {
  292. net.load_param("squeezenet_ssd.param");
  293. }
  294. void squeezenet_ssd_int8_init(ncnn::Net& net)
  295. {
  296. net.load_param("squeezenet_ssd_int8.param");
  297. }
  298. void squeezenet_ssd_run(const ncnn::Net& net)
  299. {
  300. ncnn::Extractor ex = net.create_extractor();
  301. ncnn::Mat in(300, 300, 3);
  302. ex.input("data", in);
  303. ncnn::Mat out;
  304. ex.extract("detection_out", out);
  305. }
  306. void mobilenet_ssd_init(ncnn::Net& net)
  307. {
  308. net.load_param("mobilenet_ssd.param");
  309. }
  310. void mobilenet_ssd_int8_init(ncnn::Net& net)
  311. {
  312. net.load_param("mobilenet_ssd_int8.param");
  313. }
  314. void mobilenet_ssd_run(const ncnn::Net& net)
  315. {
  316. ncnn::Extractor ex = net.create_extractor();
  317. ncnn::Mat in(300, 300, 3);
  318. ex.input("data", in);
  319. ncnn::Mat out;
  320. ex.extract("detection_out", out);
  321. }
  322. void mobilenet_yolo_init(ncnn::Net& net)
  323. {
  324. net.load_param("mobilenet_yolo.param");
  325. }
  326. void mobilenet_yolo_run(const ncnn::Net& net)
  327. {
  328. ncnn::Extractor ex = net.create_extractor();
  329. ncnn::Mat in(416, 416, 3);
  330. ex.input("data", in);
  331. ncnn::Mat out;
  332. ex.extract("detection_out", out);
  333. }
  334. void mobilenet_yolov3_init(ncnn::Net& net)
  335. {
  336. net.load_param("mobilenet_yolov3.param");
  337. }
  338. void mobilenet_yolov3_run(const ncnn::Net& net)
  339. {
  340. ncnn::Extractor ex = net.create_extractor();
  341. ncnn::Mat in(416, 416, 3);
  342. ex.input("data", in);
  343. ncnn::Mat out;
  344. ex.extract("detection_out", out);
  345. }
  346. int main(int argc, char** argv)
  347. {
  348. int loop_count = 4;
  349. int num_threads = ncnn::get_cpu_count();
  350. int powersave = 0;
  351. int gpu_device = -1;
  352. if (argc >= 2)
  353. {
  354. loop_count = atoi(argv[1]);
  355. }
  356. if (argc >= 3)
  357. {
  358. num_threads = atoi(argv[2]);
  359. }
  360. if (argc >= 4)
  361. {
  362. powersave = atoi(argv[3]);
  363. }
  364. if (argc >= 5)
  365. {
  366. gpu_device = atoi(argv[4]);
  367. }
  368. g_loop_count = loop_count;
  369. g_blob_pool_allocator.set_size_compare_ratio(0.0f);
  370. g_workspace_pool_allocator.set_size_compare_ratio(0.5f);
  371. #if NCNN_VULKAN
  372. g_use_vulkan_compute = gpu_device != -1;
  373. if (g_use_vulkan_compute)
  374. {
  375. g_vkdev = new ncnn::VulkanDevice(gpu_device);
  376. g_blob_vkallocator = new ncnn::VkUnlockedBlobBufferAllocator(g_vkdev);
  377. g_staging_vkallocator = new ncnn::VkUnlockedStagingBufferAllocator(g_vkdev);
  378. }
  379. #endif // NCNN_VULKAN
  380. ncnn::Option opt;
  381. opt.lightmode = true;
  382. opt.num_threads = num_threads;
  383. opt.blob_allocator = &g_blob_pool_allocator;
  384. opt.workspace_allocator = &g_workspace_pool_allocator;
  385. #if NCNN_VULKAN
  386. opt.vulkan_compute = g_use_vulkan_compute;
  387. opt.blob_vkallocator = g_blob_vkallocator;
  388. opt.workspace_vkallocator = g_blob_vkallocator;
  389. opt.staging_vkallocator = g_staging_vkallocator;
  390. #endif // NCNN_VULKAN
  391. ncnn::set_default_option(opt);
  392. ncnn::set_cpu_powersave(powersave);
  393. ncnn::set_omp_dynamic(0);
  394. ncnn::set_omp_num_threads(num_threads);
  395. fprintf(stderr, "loop_count = %d\n", g_loop_count);
  396. fprintf(stderr, "num_threads = %d\n", num_threads);
  397. fprintf(stderr, "powersave = %d\n", ncnn::get_cpu_powersave());
  398. fprintf(stderr, "gpu_device = %d\n", gpu_device);
  399. // run
  400. benchmark("squeezenet", squeezenet_init, squeezenet_run);
  401. #if NCNN_VULKAN
  402. if (!g_use_vulkan_compute)
  403. #endif // NCNN_VULKAN
  404. benchmark("squeezenet-int8", squeezenet_int8_init, squeezenet_run);
  405. benchmark("mobilenet", mobilenet_init, mobilenet_run);
  406. #if NCNN_VULKAN
  407. if (!g_use_vulkan_compute)
  408. #endif // NCNN_VULKAN
  409. benchmark("mobilenet-int8", mobilenet_int8_init, mobilenet_run);
  410. benchmark("mobilenet_v2", mobilenet_v2_init, mobilenet_v2_run);
  411. // #if NCNN_VULKAN
  412. // if (!g_use_vulkan_compute)
  413. // #endif // NCNN_VULKAN
  414. // benchmark("mobilenet_v2-int8", mobilenet_v2_int8_init, mobilenet_v2_run);
  415. benchmark("shufflenet", shufflenet_init, shufflenet_run);
  416. benchmark("mnasnet", mnasnet_init, mnasnet_run);
  417. benchmark("proxylessnasnet", proxylessnasnet_init, proxylessnasnet_run);
  418. benchmark("googlenet", googlenet_init, googlenet_run);
  419. #if NCNN_VULKAN
  420. if (!g_use_vulkan_compute)
  421. #endif // NCNN_VULKAN
  422. benchmark("googlenet-int8", googlenet_int8_init, googlenet_run);
  423. benchmark("resnet18", resnet18_init, resnet18_run);
  424. #if NCNN_VULKAN
  425. if (!g_use_vulkan_compute)
  426. #endif // NCNN_VULKAN
  427. benchmark("resnet18-int8", resnet18_int8_init, resnet18_run);
  428. benchmark("alexnet", alexnet_init, alexnet_run);
  429. benchmark("vgg16", vgg16_init, vgg16_run);
  430. benchmark("resnet50", resnet50_init, resnet50_run);
  431. #if NCNN_VULKAN
  432. if (!g_use_vulkan_compute)
  433. #endif // NCNN_VULKAN
  434. benchmark("resnet50-int8", resnet50_int8_init, resnet50_run);
  435. benchmark("squeezenet-ssd", squeezenet_ssd_init, squeezenet_ssd_run);
  436. #if NCNN_VULKAN
  437. if (!g_use_vulkan_compute)
  438. #endif // NCNN_VULKAN
  439. benchmark("squeezenet-ssd-int8", squeezenet_ssd_int8_init, squeezenet_ssd_run);
  440. benchmark("mobilenet-ssd", mobilenet_ssd_init, mobilenet_ssd_run);
  441. #if NCNN_VULKAN
  442. if (!g_use_vulkan_compute)
  443. #endif // NCNN_VULKAN
  444. benchmark("mobilenet-ssd-int8", mobilenet_ssd_int8_init, mobilenet_ssd_run);
  445. benchmark("mobilenet-yolo", mobilenet_yolo_init, mobilenet_yolo_run);
  446. benchmark("mobilenet-yolov3", mobilenet_yolov3_init, mobilenet_yolov3_run);
  447. #if NCNN_VULKAN
  448. delete g_blob_vkallocator;
  449. delete g_staging_vkallocator;
  450. delete g_vkdev;
  451. #endif // NCNN_VULKAN
  452. return 0;
  453. }