 [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 |
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319 |
- // Tencent is pleased to support the open source community by making ncnn available.
- //
- // Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
- //
- // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
- // in compliance with the License. You may obtain a copy of the License at
- //
- // https://opensource.org/licenses/BSD-3-Clause
- //
- // Unless required by applicable law or agreed to in writing, software distributed
- // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
- // CONDITIONS OF ANY KIND, either express or implied. See the License for the
- // specific language governing permissions and limitations under the License.
-
- #include "pooling.h"
- #include <float.h>
- #include <algorithm>
- #include "layer_type.h"
-
- namespace ncnn {
-
- DEFINE_LAYER_CREATOR(Pooling)
-
- Pooling::Pooling()
- {
- one_blob_only = true;
- support_inplace = false;
- }
-
- int Pooling::load_param(const ParamDict& pd)
- {
- pooling_type = pd.get(0, 0);
- kernel_w = pd.get(1, 0);
- kernel_h = pd.get(11, kernel_w);
- stride_w = pd.get(2, 1);
- stride_h = pd.get(12, stride_w);
- pad_left = pd.get(3, 0);
- pad_right = pd.get(14, pad_left);
- pad_top = pd.get(13, pad_left);
- pad_bottom = pd.get(15, pad_top);
- global_pooling = pd.get(4, 0);
- pad_mode = pd.get(5, 0);
- avgpool_count_include_pad = pd.get(6, 0);
-
- return 0;
- }
-
- int Pooling::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
- {
- // max value in NxN window
- // avg value in NxN window
-
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int channels = bottom_blob.c;
- size_t elemsize = bottom_blob.elemsize;
-
- // fprintf(stderr, "Pooling input %d x %d pad = %d %d %d %d ksize=%d %d stride=%d %d\n", w, h, pad_left, pad_right, pad_top, pad_bottom, kernel_w, kernel_h, stride_w, stride_h);
- if (global_pooling)
- {
- top_blob.create(channels, elemsize, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
-
- int size = w * h;
-
- if (pooling_type == PoolMethod_MAX)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q=0; q<channels; q++)
- {
- const float* ptr = bottom_blob.channel(q);
-
- float max = ptr[0];
- for (int i=0; i<size; i++)
- {
- max = std::max(max, ptr[i]);
- }
-
- top_blob[q] = max;
- }
- }
- else if (pooling_type == PoolMethod_AVE)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q=0; q<channels; q++)
- {
- const float* ptr = bottom_blob.channel(q);
-
- float sum = 0.f;
- for (int i=0; i<size; i++)
- {
- sum += ptr[i];
- }
-
- top_blob[q] = sum / size;
- }
- }
-
- return 0;
- }
-
- Mat bottom_blob_bordered = bottom_blob;
-
- float pad_value = 0.f;
- if (pooling_type == PoolMethod_MAX)
- {
- pad_value = -FLT_MAX;
- }
- else if (pooling_type == PoolMethod_AVE)
- {
- pad_value = 0.f;
- }
-
- int wtailpad = 0;
- int htailpad = 0;
-
- if (pad_mode == 0) // full padding
- {
- int wtail = (w + pad_left + pad_right - kernel_w) % stride_w;
- int htail = (h + pad_top + pad_bottom - kernel_h) % stride_h;
-
- if (wtail != 0)
- wtailpad = stride_w - wtail;
- if (htail != 0)
- htailpad = stride_h - htail;
-
- Option opt_b = opt;
- opt_b.blob_allocator = opt.workspace_allocator;
- copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom + htailpad, pad_left, pad_right + wtailpad, BORDER_CONSTANT, pad_value, opt_b);
- if (bottom_blob_bordered.empty())
- return -100;
-
- w = bottom_blob_bordered.w;
- h = bottom_blob_bordered.h;
- }
- else if (pad_mode == 1) // valid padding
- {
- Option opt_b = opt;
- opt_b.blob_allocator = opt.workspace_allocator;
- copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, pad_value, opt_b);
- if (bottom_blob_bordered.empty())
- return -100;
-
- w = bottom_blob_bordered.w;
- h = bottom_blob_bordered.h;
- }
- else if (pad_mode == 2) // tensorflow padding=SAME or onnx padding=SAME_UPPER
- {
- int wpad = kernel_w + (w - 1) / stride_w * stride_w - w;
- int hpad = kernel_h + (h - 1) / stride_h * stride_h - h;
- if (wpad > 0 || hpad > 0)
- {
- Option opt_b = opt;
- opt_b.blob_allocator = opt.workspace_allocator;
- copy_make_border(bottom_blob, bottom_blob_bordered, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, BORDER_CONSTANT, pad_value, opt_b);
- if (bottom_blob_bordered.empty())
- return -100;
- }
-
- w = bottom_blob_bordered.w;
- h = bottom_blob_bordered.h;
- }
- else if (pad_mode == 3) // onnx padding=SAME_LOWER
- {
- int wpad = kernel_w + (w - 1) / stride_w * stride_w - w;
- int hpad = kernel_h + (h - 1) / stride_h * stride_h - h;
- if (wpad > 0 || hpad > 0)
- {
- Option opt_b = opt;
- opt_b.blob_allocator = opt.workspace_allocator;
- copy_make_border(bottom_blob, bottom_blob_bordered, hpad - hpad / 2, hpad / 2, wpad - wpad / 2, wpad / 2, BORDER_CONSTANT, pad_value, opt_b);
- if (bottom_blob_bordered.empty())
- return -100;
- }
-
- w = bottom_blob_bordered.w;
- h = bottom_blob_bordered.h;
- }
-
- int outw = (w - kernel_w) / stride_w + 1;
- int outh = (h - kernel_h) / stride_h + 1;
-
- top_blob.create(outw, outh, channels, elemsize, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
-
- const int maxk = kernel_w * kernel_h;
-
- // kernel offsets
- std::vector<int> _space_ofs(maxk);
- int* space_ofs = &_space_ofs[0];
- {
- int p1 = 0;
- int p2 = 0;
- int gap = w - kernel_w;
- for (int i = 0; i < kernel_h; i++)
- {
- for (int j = 0; j < kernel_w; j++)
- {
- space_ofs[p1] = p2;
- p1++;
- p2++;
- }
- p2 += gap;
- }
- }
-
- if (pooling_type == PoolMethod_MAX)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q=0; q<channels; q++)
- {
- const Mat m = bottom_blob_bordered.channel(q);
- float* outptr = top_blob.channel(q);
-
- for (int i = 0; i < outh; i++)
- {
- for (int j = 0; j < outw; j++)
- {
- const float* sptr = m.row(i*stride_h) + j*stride_w;
-
- float max = sptr[0];
-
- for (int k = 0; k < maxk; k++)
- {
- float val = sptr[ space_ofs[k] ];
- max = std::max(max, val);
- }
-
- outptr[j] = max;
- }
-
- outptr += outw;
- }
- }
- }
- else if (pooling_type == PoolMethod_AVE)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q=0; q<channels; q++)
- {
- const Mat m = bottom_blob_bordered.channel(q);
- float* outptr = top_blob.channel(q);
-
- for (int i = 0; i < outh; i++)
- {
- for (int j = 0; j < outw; j++)
- {
- const float* sptr = m.row(i*stride_h) + j*stride_w;
-
- float sum = 0;
-
- for (int k = 0; k < maxk; k++)
- {
- float val = sptr[ space_ofs[k] ];
- sum += val;
- }
-
- outptr[j] = sum / maxk;
- }
-
- outptr += outw;
- }
-
- if (avgpool_count_include_pad == 0)
- {
- // fix pad
- if (pad_top != 0)
- {
- const float scale = (float)kernel_h / (kernel_h - pad_top);
-
- outptr = top_blob.channel(q).row(0);
- for (int i = 0; i < outw; i++)
- {
- outptr[i] *= scale;
- }
- }
- if (pad_bottom + htailpad != 0)
- {
- const float scale = (float)kernel_h / (kernel_h - pad_bottom - htailpad);
-
- outptr = top_blob.channel(q).row(outh - 1);
- for (int i = 0; i < outw; i++)
- {
- outptr[i] *= scale;
- }
- }
- if (pad_left != 0)
- {
- const float scale = (float)kernel_w / (kernel_w - pad_left);
-
- outptr = top_blob.channel(q);
- for (int i = 0; i < outh; i++)
- {
- *outptr *= scale;
- outptr += outw;
- }
- }
- if (pad_right + wtailpad != 0)
- {
- const float scale = (float)kernel_w / (kernel_w - pad_right - wtailpad);
-
- outptr = top_blob.channel(q);
- outptr += outw - 1;
- for (int i = 0; i < outh; i++)
- {
- *outptr *= scale;
- outptr += outw;
- }
- }
- }
- }
- }
-
- return 0;
- }
-
- } // namespace ncnn
|