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.

pipeline.cpp 17 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
[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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561
  1. // Copyright 2019 Tencent
  2. // SPDX-License-Identifier: BSD-3-Clause
  3. #include "pipeline.h"
  4. #include "layer_shader_type.h"
  5. #include "mat.h"
  6. #include "pipelinecache.h"
  7. #include "option.h"
  8. #if __ANDROID_API__ >= 26
  9. #include <android/hardware_buffer.h>
  10. #endif // __ANDROID_API__ >= 26
  11. namespace ncnn {
  12. #if NCNN_VULKAN
  13. class PipelinePrivate
  14. {
  15. public:
  16. VkShaderModule shader_module;
  17. VkDescriptorSetLayout descriptorset_layout;
  18. VkPipelineLayout pipeline_layout;
  19. VkPipeline pipeline;
  20. VkDescriptorUpdateTemplateKHR descriptor_update_template;
  21. ShaderInfo shader_info;
  22. uint32_t local_size_x;
  23. uint32_t local_size_y;
  24. uint32_t local_size_z;
  25. uint32_t subgroup_size;
  26. };
  27. Pipeline::Pipeline(const VulkanDevice* _vkdev)
  28. : vkdev(_vkdev), d(new PipelinePrivate)
  29. {
  30. d->shader_module = 0;
  31. d->descriptorset_layout = 0;
  32. d->pipeline_layout = 0;
  33. d->pipeline = 0;
  34. d->descriptor_update_template = 0;
  35. d->local_size_x = 1;
  36. d->local_size_y = 1;
  37. d->local_size_z = 1;
  38. d->subgroup_size = vkdev->info.subgroup_size();
  39. }
  40. Pipeline::~Pipeline()
  41. {
  42. delete d;
  43. }
  44. Pipeline::Pipeline(const Pipeline&)
  45. : d(0)
  46. {
  47. }
  48. Pipeline& Pipeline::operator=(const Pipeline&)
  49. {
  50. return *this;
  51. }
  52. void Pipeline::set_optimal_local_size_xyz(int w, int h, int c)
  53. {
  54. set_optimal_local_size_xyz(Mat(w, h, c, (void*)0));
  55. }
  56. void Pipeline::set_optimal_local_size_xyz(const Mat& local_size_xyz)
  57. {
  58. int w = local_size_xyz.w;
  59. int h = local_size_xyz.h;
  60. int c = local_size_xyz.c;
  61. if (w == 0 && h == 0 && c == 0)
  62. {
  63. // fallback to the common and safe 4x4x4
  64. w = 4;
  65. h = 4;
  66. c = 4;
  67. }
  68. w = std::min(w, (int)vkdev->info.max_workgroup_size_x());
  69. h = std::min(h, (int)vkdev->info.max_workgroup_size_y());
  70. c = std::min(c, (int)vkdev->info.max_workgroup_size_z());
  71. if (w * h * c <= (int)vkdev->info.max_workgroup_invocations())
  72. {
  73. return set_local_size_xyz(w, h, c);
  74. }
  75. int max_local_size_xy = (int)vkdev->info.max_workgroup_invocations() / c;
  76. int wh_max = std::max(1, (int)sqrt(max_local_size_xy));
  77. while (w * h >= wh_max)
  78. {
  79. w = std::max(1, w / 2);
  80. h = std::max(1, h / 2);
  81. }
  82. set_local_size_xyz(w, h, c);
  83. }
  84. void Pipeline::set_subgroup_size(uint32_t subgroup_size)
  85. {
  86. // assert subgroup_size be power of two
  87. subgroup_size = std::max(subgroup_size, vkdev->info.min_subgroup_size());
  88. subgroup_size = std::min(subgroup_size, vkdev->info.max_subgroup_size());
  89. d->subgroup_size = subgroup_size;
  90. }
  91. static int count_trailing_zeros(unsigned int v)
  92. {
  93. int cnt = 0;
  94. while ((v & 1) == 0)
  95. {
  96. cnt++;
  97. v >>= 1;
  98. }
  99. return cnt;
  100. }
  101. // round up v to the next multiple of 2^k
  102. static unsigned int round_up_pow2_mul(unsigned int v, int k)
  103. {
  104. unsigned int m = 1u << k;
  105. return ((v + m - 1) / m) * m;
  106. }
  107. // adjust x, y, z so that new x * y * z is a multiple of size (size must be a power of two), and new x, y, z are no less than the inputs
  108. // new values do not have to be integer multiples of the originals
  109. // minimize the total increment (x'-x)+(y'-y)+(z'-z)
  110. // additional constraint: if original y is 1, prefer not to adjust y; if original z is 1, prefer not to adjust z
  111. static void adjust_xyz(int& x, int& y, int& z, const int subgroup_size)
  112. {
  113. if (x * y * z % subgroup_size == 0)
  114. return;
  115. int target_n = 0;
  116. {
  117. while ((1 << target_n) != subgroup_size)
  118. target_n++;
  119. }
  120. // subgroup shall usually be 4 ~ 128, sanitize the max possible size
  121. target_n = std::min(target_n, 10);
  122. const int tx = count_trailing_zeros((unsigned int)x);
  123. const int ty = count_trailing_zeros((unsigned int)y);
  124. const int tz = count_trailing_zeros((unsigned int)z);
  125. const int tn = tx + ty + tz;
  126. const int need = target_n - tn;
  127. if (z == 1)
  128. {
  129. if (y == 1)
  130. {
  131. // adjust x only
  132. x = round_up_pow2_mul((unsigned int)x, target_n);
  133. }
  134. else if (x == 1)
  135. {
  136. // adjust y only
  137. y = round_up_pow2_mul((unsigned int)y, target_n);
  138. }
  139. else
  140. {
  141. // adjust x and y
  142. y = round_up_pow2_mul((unsigned int)y, ty + need / 2);
  143. x = round_up_pow2_mul((unsigned int)x, tx + need - need / 2);
  144. }
  145. }
  146. else if (y == 1)
  147. {
  148. if (x == 1)
  149. {
  150. // adjust z only
  151. z = round_up_pow2_mul((unsigned int)z, target_n);
  152. }
  153. else
  154. {
  155. // adjust x and z
  156. z = round_up_pow2_mul((unsigned int)z, tz + need / 2);
  157. x = round_up_pow2_mul((unsigned int)x, tx + need - need / 2);
  158. }
  159. }
  160. else if (x == 1)
  161. {
  162. // adjust y and z
  163. z = round_up_pow2_mul((unsigned int)z, tz + need / 2);
  164. y = round_up_pow2_mul((unsigned int)y, ty + need - need / 2);
  165. }
  166. else
  167. {
  168. // adjust x y z
  169. z = round_up_pow2_mul((unsigned int)z, tz + need / 3);
  170. y = round_up_pow2_mul((unsigned int)y, ty + (need - need / 3) / 2);
  171. x = round_up_pow2_mul((unsigned int)x, tx + need - (need - need / 3) / 2);
  172. }
  173. }
  174. void Pipeline::set_local_size_xyz(int w, int h, int c)
  175. {
  176. // dispatch at least one subgroup
  177. // make local size be multiple of subgroup size
  178. // and metal is unhappy with arbitrary local size anyway
  179. adjust_xyz(w, h, c, d->subgroup_size);
  180. d->local_size_x = w;
  181. d->local_size_y = h;
  182. d->local_size_z = c;
  183. // NCNN_LOGE("local size = %d %d %d", local_size_x, local_size_y, local_size_z);
  184. }
  185. int Pipeline::create(const uint32_t* spv_data, size_t spv_data_size, const std::vector<vk_specialization_type>& specializations)
  186. {
  187. const PipelineCache* pipeline_cache = vkdev->get_pipeline_cache();
  188. // get from pipeline cache
  189. return pipeline_cache->get_pipeline(spv_data, spv_data_size, specializations, d->local_size_x, d->local_size_y, d->local_size_z, d->subgroup_size,
  190. &d->shader_module, &d->descriptorset_layout, &d->pipeline_layout, &d->pipeline, &d->descriptor_update_template,
  191. d->shader_info);
  192. }
  193. int Pipeline::create(int shader_type_index, const Option& opt, const std::vector<vk_specialization_type>& specializations)
  194. {
  195. const PipelineCache* pipeline_cache = opt.pipeline_cache ? opt.pipeline_cache : vkdev->get_pipeline_cache();
  196. // get from pipeline cache
  197. return pipeline_cache->get_pipeline(shader_type_index, opt, specializations, d->local_size_x, d->local_size_y, d->local_size_z, d->subgroup_size,
  198. &d->shader_module, &d->descriptorset_layout, &d->pipeline_layout, &d->pipeline, &d->descriptor_update_template,
  199. d->shader_info);
  200. }
  201. VkShaderModule Pipeline::shader_module() const
  202. {
  203. return d->shader_module;
  204. }
  205. VkDescriptorSetLayout Pipeline::descriptorset_layout() const
  206. {
  207. return d->descriptorset_layout;
  208. }
  209. VkPipelineLayout Pipeline::pipeline_layout() const
  210. {
  211. return d->pipeline_layout;
  212. }
  213. VkPipeline Pipeline::pipeline() const
  214. {
  215. return d->pipeline;
  216. }
  217. VkDescriptorUpdateTemplateKHR Pipeline::descriptor_update_template() const
  218. {
  219. return d->descriptor_update_template;
  220. }
  221. const ShaderInfo& Pipeline::shader_info() const
  222. {
  223. return d->shader_info;
  224. }
  225. uint32_t Pipeline::local_size_x() const
  226. {
  227. return d->local_size_x;
  228. }
  229. uint32_t Pipeline::local_size_y() const
  230. {
  231. return d->local_size_y;
  232. }
  233. uint32_t Pipeline::local_size_z() const
  234. {
  235. return d->local_size_z;
  236. }
  237. void Pipeline::set_shader_module(VkShaderModule shader_module)
  238. {
  239. d->shader_module = shader_module;
  240. }
  241. void Pipeline::set_descriptorset_layout(VkDescriptorSetLayout descriptorset_layout)
  242. {
  243. d->descriptorset_layout = descriptorset_layout;
  244. }
  245. void Pipeline::set_pipeline_layout(VkPipelineLayout pipeline_layout)
  246. {
  247. d->pipeline_layout = pipeline_layout;
  248. }
  249. void Pipeline::set_pipeline(VkPipeline pipeline)
  250. {
  251. d->pipeline = pipeline;
  252. }
  253. void Pipeline::set_descriptor_update_template(VkDescriptorUpdateTemplateKHR descriptor_update_template)
  254. {
  255. d->descriptor_update_template = descriptor_update_template;
  256. }
  257. void Pipeline::set_shader_info(const ShaderInfo& shader_info)
  258. {
  259. d->shader_info = shader_info;
  260. }
  261. #if NCNN_PLATFORM_API
  262. #if __ANDROID_API__ >= 26
  263. ImportAndroidHardwareBufferPipeline::ImportAndroidHardwareBufferPipeline(const VulkanDevice* _vkdev)
  264. : Pipeline(_vkdev)
  265. {
  266. sampler = 0;
  267. }
  268. ImportAndroidHardwareBufferPipeline::~ImportAndroidHardwareBufferPipeline()
  269. {
  270. destroy();
  271. }
  272. int ImportAndroidHardwareBufferPipeline::create(VkAndroidHardwareBufferImageAllocator* ahb_im_allocator, int _type_to, int _rotate_from, const Option& opt)
  273. {
  274. int target_width;
  275. int target_height;
  276. if (rotate_from < 5) // 1 2 3 4
  277. {
  278. target_width = ahb_im_allocator->width();
  279. target_height = ahb_im_allocator->height();
  280. }
  281. else // 5 6 7 8
  282. {
  283. target_width = ahb_im_allocator->height();
  284. target_height = ahb_im_allocator->width();
  285. }
  286. return create(ahb_im_allocator, _type_to, _rotate_from, target_width, target_height, opt);
  287. }
  288. int ImportAndroidHardwareBufferPipeline::create(VkAndroidHardwareBufferImageAllocator* ahb_im_allocator, int _type_to, int _rotate_from, int target_width, int target_height, const Option& opt)
  289. {
  290. int w = ahb_im_allocator->width();
  291. int h = ahb_im_allocator->height();
  292. type_to = _type_to;
  293. rotate_from = _rotate_from;
  294. need_resize = false;
  295. if (rotate_from < 5) // 1 2 3 4
  296. {
  297. if (target_width != w || target_height != h)
  298. need_resize = true;
  299. }
  300. else // 5 6 7 8
  301. {
  302. if (target_width != h || target_height != w)
  303. need_resize = true;
  304. }
  305. // if (type_to == 1 || type_to == 2)
  306. // {
  307. // outc = 3;
  308. // out_elemsize = vkdev->info.support_fp16_storage() && opt.use_fp16_storage ? 2u : 4u;
  309. // out_elempack = 1;
  310. // }
  311. // else if (type_to == 3)
  312. // {
  313. // outc = 1;
  314. // out_elemsize = vkdev->info.support_fp16_storage() && opt.use_fp16_storage ? 2u : 4u;
  315. // out_elempack = 1;
  316. // }
  317. // else // if (type_to == 4 || type_to == 5)
  318. // {
  319. // outc = 1;
  320. // out_elemsize = ((vkdev->info.support_fp16_packed() && opt.use_fp16_packed) || (vkdev->info.support_fp16_storage() && opt.use_fp16_storage)) ? 8u : 16u;
  321. // out_elempack = 4;
  322. // }
  323. set_local_size_xyz(8, 8, 1);
  324. std::vector<vk_specialization_type> specializations(7);
  325. specializations[0].i = ahb_im_allocator->width();
  326. specializations[1].i = ahb_im_allocator->height();
  327. specializations[2].i = target_width;
  328. specializations[3].i = target_height;
  329. specializations[4].i = type_to;
  330. specializations[5].i = rotate_from;
  331. specializations[6].i = need_resize;
  332. create_shader_module(opt);
  333. const ShaderInfo& _shader_info = shader_info();
  334. if ((int)specializations.size() != _shader_info.specialization_count)
  335. {
  336. NCNN_LOGE("pipeline convert_ycbcr specialization count mismatch, expect %d but got %d", _shader_info.specialization_count, (int)specializations.size());
  337. return -1;
  338. }
  339. create_sampler(ahb_im_allocator);
  340. create_descriptorset_layout();
  341. VkPipelineLayout pipeline_layout = 0;
  342. VkPipeline pipeline = 0;
  343. VkDescriptorUpdateTemplateKHR descriptor_update_template = 0;
  344. vkdev->create_pipeline_layout(_shader_info.push_constant_count, descriptorset_layout(), &pipeline_layout);
  345. vkdev->create_pipeline(shader_module(), pipeline_layout, specializations, vkdev->info.subgroup_size(), &pipeline);
  346. if (vkdev->info.support_VK_KHR_descriptor_update_template())
  347. {
  348. vkdev->create_descriptor_update_template(_shader_info.binding_count, _shader_info.binding_types, descriptorset_layout(), pipeline_layout, &descriptor_update_template);
  349. }
  350. set_pipeline_layout(pipeline_layout);
  351. set_pipeline(pipeline);
  352. set_descriptor_update_template(descriptor_update_template);
  353. return 0;
  354. }
  355. void ImportAndroidHardwareBufferPipeline::destroy()
  356. {
  357. if (sampler)
  358. {
  359. vkDestroySampler(vkdev->vkdevice(), sampler, 0);
  360. sampler = 0;
  361. }
  362. }
  363. int ImportAndroidHardwareBufferPipeline::create_shader_module(const Option& opt)
  364. {
  365. int shader_type_index = LayerShaderType::convert_ycbcr;
  366. std::vector<uint32_t> spirv;
  367. int retc = compile_spirv_module(shader_type_index, opt, spirv);
  368. if (retc != 0)
  369. {
  370. NCNN_LOGE("compile_spirv_module failed %d", retc);
  371. return -1;
  372. }
  373. const uint32_t* spv_data = spirv.data();
  374. size_t spv_data_size = spirv.size() * 4;
  375. ShaderInfo shader_info;
  376. int ret = resolve_shader_info(spv_data, spv_data_size, shader_info);
  377. if (ret != 0)
  378. {
  379. NCNN_LOGE("resolve_shader_info failed %d", ret);
  380. return -1;
  381. }
  382. set_shader_info(shader_info);
  383. VkShaderModule shader_module = vkdev->compile_shader_module(spv_data, spv_data_size, local_size_x(), local_size_y(), local_size_z());
  384. set_shader_module(shader_module);
  385. return 0;
  386. }
  387. int ImportAndroidHardwareBufferPipeline::create_sampler(VkAndroidHardwareBufferImageAllocator* ahb_im_allocator)
  388. {
  389. VkResult ret;
  390. VkExternalFormatANDROID externalFormatANDROID;
  391. externalFormatANDROID.sType = VK_STRUCTURE_TYPE_EXTERNAL_FORMAT_ANDROID;
  392. externalFormatANDROID.pNext = 0;
  393. externalFormatANDROID.externalFormat = ahb_im_allocator->external_format();
  394. VkSamplerYcbcrConversionInfoKHR samplerYcbcrConversionInfo;
  395. samplerYcbcrConversionInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_INFO_KHR;
  396. samplerYcbcrConversionInfo.pNext = &externalFormatANDROID;
  397. samplerYcbcrConversionInfo.conversion = ahb_im_allocator->samplerYcbcrConversion;
  398. VkSamplerCreateInfo samplerCreateInfo;
  399. samplerCreateInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
  400. samplerCreateInfo.pNext = &samplerYcbcrConversionInfo;
  401. samplerCreateInfo.magFilter = need_resize ? VK_FILTER_LINEAR : VK_FILTER_NEAREST;
  402. samplerCreateInfo.minFilter = need_resize ? VK_FILTER_LINEAR : VK_FILTER_NEAREST;
  403. samplerCreateInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST;
  404. samplerCreateInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
  405. samplerCreateInfo.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
  406. samplerCreateInfo.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
  407. samplerCreateInfo.mipLodBias = 0.0f;
  408. samplerCreateInfo.anisotropyEnable = VK_FALSE;
  409. samplerCreateInfo.maxAnisotropy = 1;
  410. samplerCreateInfo.compareEnable = VK_FALSE;
  411. samplerCreateInfo.compareOp = VK_COMPARE_OP_NEVER;
  412. samplerCreateInfo.minLod = 0.0f;
  413. samplerCreateInfo.maxLod = 0.0f;
  414. samplerCreateInfo.borderColor = VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK; //VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE; FIXME
  415. samplerCreateInfo.unnormalizedCoordinates = VK_TRUE; //VK_FALSE; FIXME ?
  416. ret = vkCreateSampler(vkdev->vkdevice(), &samplerCreateInfo, 0, &sampler);
  417. if (ret != VK_SUCCESS)
  418. {
  419. NCNN_LOGE("vkCreateSampler failed %d", ret);
  420. return -1;
  421. }
  422. return 0;
  423. }
  424. int ImportAndroidHardwareBufferPipeline::create_descriptorset_layout()
  425. {
  426. VkDescriptorSetLayoutBinding descriptorSetLayoutBindings[3];
  427. descriptorSetLayoutBindings[0].binding = 0;
  428. descriptorSetLayoutBindings[0].descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
  429. descriptorSetLayoutBindings[0].descriptorCount = 1;
  430. descriptorSetLayoutBindings[0].stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
  431. descriptorSetLayoutBindings[0].pImmutableSamplers = &sampler;
  432. descriptorSetLayoutBindings[1].binding = 1;
  433. descriptorSetLayoutBindings[1].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
  434. descriptorSetLayoutBindings[1].descriptorCount = 1;
  435. descriptorSetLayoutBindings[1].stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
  436. descriptorSetLayoutBindings[1].pImmutableSamplers = 0;
  437. descriptorSetLayoutBindings[2].binding = 2;
  438. descriptorSetLayoutBindings[2].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
  439. descriptorSetLayoutBindings[2].descriptorCount = 1;
  440. descriptorSetLayoutBindings[2].stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
  441. descriptorSetLayoutBindings[2].pImmutableSamplers = 0;
  442. VkDescriptorSetLayoutCreateInfo descriptorSetLayoutCreateInfo;
  443. descriptorSetLayoutCreateInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
  444. descriptorSetLayoutCreateInfo.pNext = 0;
  445. descriptorSetLayoutCreateInfo.flags = 0;
  446. descriptorSetLayoutCreateInfo.bindingCount = 3;
  447. descriptorSetLayoutCreateInfo.pBindings = descriptorSetLayoutBindings;
  448. if (vkdev->info.support_VK_KHR_push_descriptor())
  449. {
  450. descriptorSetLayoutCreateInfo.flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR;
  451. }
  452. VkDescriptorSetLayout descriptorset_layout = 0;
  453. VkResult ret = vkCreateDescriptorSetLayout(vkdev->vkdevice(), &descriptorSetLayoutCreateInfo, 0, &descriptorset_layout);
  454. if (ret != VK_SUCCESS)
  455. {
  456. NCNN_LOGE("vkCreateDescriptorSetLayout failed %d", ret);
  457. return -1;
  458. }
  459. set_descriptorset_layout(descriptorset_layout);
  460. return 0;
  461. }
  462. #endif // __ANDROID_API__ >= 26
  463. #endif // NCNN_PLATFORM_API
  464. #endif // NCNN_VULKAN
  465. } // namespace ncnn