git.sesse.net Git - casparcg/blob - accelerator/ogl/util/device.cpp

   1 /*
   2 * Copyright (c) 2011 Sveriges Television AB <info@casparcg.com>
   3 *
   4 * This file is part of CasparCG (www.casparcg.com).
   5 *
   6 * CasparCG is free software: you can redistribute it and/or modify
   7 * it under the terms of the GNU General Public License as published by
   8 * the Free Software Foundation, either version 3 of the License, or
   9 * (at your option) any later version.
  10 *
  11 * CasparCG is distributed in the hope that it will be useful,
  12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 * GNU General Public License for more details.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * along with CasparCG. If not, see <http://www.gnu.org/licenses/>.
  18 *
  19 * Author: Robert Nagy, ronag89@gmail.com
  20 */
  21
  22 // TODO: Smart GC
  23
  24 #include "../../StdAfx.h"
  25
  26 #include "device.h"
  27
  28 #include "buffer.h"
  29 #include "texture.h"
  30 #include "shader.h"
  31
  32 #include <common/assert.h>
  33 #include <common/except.h>
  34 #include <common/future.h>
  35 #include <common/array.h>
  36 #include <common/memory.h>
  37 #include <common/gl/gl_check.h>
  38 #include <common/timer.h>
  39
  40 #include <GL/glew.h>
  41
  42 #include <SFML/Window/Context.hpp>
  43
  44 #include <tbb/concurrent_unordered_map.h>
  45 #include <tbb/concurrent_hash_map.h>
  46 #include <tbb/concurrent_queue.h>
  47
  48 #include <boost/utility/declval.hpp>
  49 #include <boost/property_tree/ptree.hpp>
  50
  51 #include <array>
  52 #include <unordered_map>
  53
  54 #include <asmlib.h>
  55 #include <tbb/parallel_for.h>
  56
  57 namespace caspar { namespace accelerator { namespace ogl {
  58
  59 struct device::impl : public std::enable_shared_from_this<impl>
  60 {
  61         static_assert(std::is_same<decltype(boost::declval<device>().impl_), spl::shared_ptr<impl>>::value, "impl_ must be shared_ptr");
  62
  63         tbb::concurrent_hash_map<buffer*, std::shared_ptr<texture>> texture_cache_;
  64
  65         std::unique_ptr<sf::Context> device_;
  66
  67         std::array<tbb::concurrent_unordered_map<std::size_t, tbb::concurrent_bounded_queue<std::shared_ptr<texture>>>, 8>      device_pools_;
  68         std::array<tbb::concurrent_unordered_map<std::size_t, tbb::concurrent_bounded_queue<std::shared_ptr<buffer>>>, 2>       host_pools_;
  69
  70         GLuint fbo_;
  71
  72         executor& executor_;
  73
  74         impl(executor& executor)
  75                 : executor_(executor)
  76         {
  77                 executor_.set_capacity(256);
  78
  79                 CASPAR_LOG(info) << L"Initializing OpenGL Device.";
  80
  81                 executor_.invoke([=]
  82                 {
  83                         device_.reset(new sf::Context());
  84                         device_->setActive(true);
  85
  86                         if (glewInit() != GLEW_OK)
  87                                 CASPAR_THROW_EXCEPTION(gl::ogl_exception() << msg_info("Failed to initialize GLEW."));
  88
  89                         if(!GLEW_VERSION_3_0)
  90                                 CASPAR_THROW_EXCEPTION(not_supported() << msg_info("Your graphics card does not meet the minimum hardware requirements since it does not support OpenGL 3.0 or higher."));
  91
  92                         glGenFramebuffers(1, &fbo_);
  93                         glBindFramebuffer(GL_FRAMEBUFFER, fbo_);
  94                 });
  95
  96                 CASPAR_LOG(info) << L"Successfully initialized OpenGL " << version();
  97         }
  98
  99         ~impl()
 100         {
 101                 executor_.invoke([=]
 102                 {
 103                         texture_cache_.clear();
 104
 105                         for (auto& pool : host_pools_)
 106                                 pool.clear();
 107
 108                         for (auto& pool : device_pools_)
 109                                 pool.clear();
 110
 111                         glDeleteFramebuffers(1, &fbo_);
 112
 113                         device_.reset();
 114                 });
 115         }
 116
 117         boost::property_tree::wptree info() const
 118         {
 119                 boost::property_tree::wptree info;
 120
 121                 boost::property_tree::wptree pooled_device_buffers;
 122                 size_t total_pooled_device_buffer_size  = 0;
 123                 size_t total_pooled_device_buffer_count = 0;
 124
 125                 for (size_t i = 0; i < device_pools_.size(); ++i)
 126                 {
 127                         auto& pools             = device_pools_.at(i);
 128                         bool mipmapping = i > 3;
 129                         auto stride             = mipmapping ? i - 3 : i + 1;
 130
 131                         for (auto& pool : pools)
 132                         {
 133                                 auto width      = pool.first >> 16;
 134                                 auto height     = pool.first & 0x0000FFFF;
 135                                 auto size       = width * height * stride;
 136                                 auto count      = pool.second.size();
 137
 138                                 if (count == 0)
 139                                         continue;
 140
 141                                 boost::property_tree::wptree pool_info;
 142
 143                                 pool_info.add(L"stride",                stride);
 144                                 pool_info.add(L"mipmapping",    mipmapping);
 145                                 pool_info.add(L"width",                 width);
 146                                 pool_info.add(L"height",                height);
 147                                 pool_info.add(L"size",                  size);
 148                                 pool_info.add(L"count",                 count);
 149
 150                                 total_pooled_device_buffer_size         += size * count;
 151                                 total_pooled_device_buffer_count        += count;
 152
 153                                 pooled_device_buffers.add_child(L"device_buffer_pool", pool_info);
 154                         }
 155                 }
 156
 157                 info.add_child(L"gl.details.pooled_device_buffers", pooled_device_buffers);
 158
 159                 boost::property_tree::wptree pooled_host_buffers;
 160                 size_t total_read_size          = 0;
 161                 size_t total_write_size         = 0;
 162                 size_t total_read_count         = 0;
 163                 size_t total_write_count        = 0;
 164
 165                 for (size_t i = 0; i < host_pools_.size(); ++i)
 166                 {
 167                         auto& pools     = host_pools_.at(i);
 168                         auto usage      = static_cast<buffer::usage>(i);
 169
 170                         for (auto& pool : pools)
 171                         {
 172                                 auto size       = pool.first;
 173                                 auto count      = pool.second.size();
 174
 175                                 if (count == 0)
 176                                         continue;
 177
 178                                 boost::property_tree::wptree pool_info;
 179
 180                                 pool_info.add(L"usage", usage == buffer::usage::read_only ? L"read_only" : L"write_only");
 181                                 pool_info.add(L"size",  size);
 182                                 pool_info.add(L"count", count);
 183
 184                                 pooled_host_buffers.add_child(L"host_buffer_pool", pool_info);
 185
 186                                 (usage == buffer::usage::read_only ? total_read_count : total_write_count) += count;
 187                                 (usage == buffer::usage::read_only ? total_read_size : total_write_size) += size * count;
 188                         }
 189                 }
 190
 191                 info.add_child(L"gl.details.pooled_host_buffers",                               pooled_host_buffers);
 192                 info.add(L"gl.summary.pooled_device_buffers.total_count",               total_pooled_device_buffer_count);
 193                 info.add(L"gl.summary.pooled_device_buffers.total_size",                total_pooled_device_buffer_size);
 194                 info.add_child(L"gl.summary.all_device_buffers",                                texture::info());
 195                 info.add(L"gl.summary.pooled_host_buffers.total_read_count",    total_read_count);
 196                 info.add(L"gl.summary.pooled_host_buffers.total_write_count",   total_write_count);
 197                 info.add(L"gl.summary.pooled_host_buffers.total_read_size",             total_read_size);
 198                 info.add(L"gl.summary.pooled_host_buffers.total_write_size",    total_write_size);
 199                 info.add_child(L"gl.summary.all_host_buffers",                                  buffer::info());
 200
 201                 return info;
 202         }
 203
 204         std::wstring version()
 205         {
 206                 try
 207                 {
 208                         return executor_.invoke([]
 209                         {
 210                                 return u16(reinterpret_cast<const char*>(GL2(glGetString(GL_VERSION)))) + L" " + u16(reinterpret_cast<const char*>(GL2(glGetString(GL_VENDOR))));
 211                         });
 212                 }
 213                 catch(...)
 214                 {
 215                         return L"Not found";;
 216                 }
 217         }
 218
 219         spl::shared_ptr<texture> create_texture(int width, int height, int stride, bool mipmapped, bool clear)
 220         {
 221                 CASPAR_VERIFY(stride > 0 && stride < 5);
 222                 CASPAR_VERIFY(width > 0 && height > 0);
 223
 224                 if(!executor_.is_current())
 225                         CASPAR_THROW_EXCEPTION(invalid_operation() << msg_info("Operation only valid in an OpenGL Context."));
 226
 227                 auto pool = &device_pools_[stride - 1 + (mipmapped ? 4 : 0)][((width << 16) & 0xFFFF0000) | (height & 0x0000FFFF)];
 228
 229                 std::shared_ptr<texture> tex;
 230                 if(!pool->try_pop(tex))
 231                         tex = spl::make_shared<texture>(width, height, stride, mipmapped);
 232
 233                 if(clear)
 234                         tex->clear();
 235
 236                 return spl::shared_ptr<texture>(tex.get(), [tex, pool](texture*) mutable
 237                 {
 238                         pool->push(tex);
 239                 });
 240         }
 241
 242         spl::shared_ptr<buffer> create_buffer(std::size_t size, buffer::usage usage)
 243         {
 244                 CASPAR_VERIFY(size > 0);
 245
 246                 auto pool = &host_pools_[static_cast<int>(usage)][size];
 247
 248                 std::shared_ptr<buffer> buf;
 249                 if(!pool->try_pop(buf))
 250                 {
 251                         caspar::timer timer;
 252
 253                         buf = executor_.invoke([&]
 254                         {
 255                                 return std::make_shared<buffer>(size, usage);
 256                         }, task_priority::high_priority);
 257
 258                         if(timer.elapsed() > 0.02)
 259                                 CASPAR_LOG(warning) << L"[ogl-device] Performance warning. Buffer allocation blocked: " << timer.elapsed();
 260                 }
 261
 262                 std::weak_ptr<impl> self = shared_from_this(); // buffers can leave the device context, take a hold on life-time.
 263                 return spl::shared_ptr<buffer>(buf.get(), [=](buffer*) mutable
 264                 {
 265                         auto strong = self.lock();
 266
 267                         if (strong)
 268                         {
 269                                 strong->texture_cache_.erase(buf.get());
 270                                 pool->push(buf);
 271                         }
 272                         else
 273                         {
 274                                 CASPAR_LOG(info) << L"Buffer outlived ogl device";
 275                         }
 276                 });
 277         }
 278
 279         array<std::uint8_t> create_array(std::size_t size)
 280         {
 281                 auto buf = create_buffer(size, buffer::usage::write_only);
 282                 return array<std::uint8_t>(buf->data(), buf->size(), false, buf);
 283         }
 284
 285         template<typename T>
 286         std::shared_ptr<buffer> copy_to_buf(const T& source)
 287         {
 288                 std::shared_ptr<buffer> buf;
 289
 290                 auto tmp = source.template storage<spl::shared_ptr<buffer>>();
 291                 if(tmp)
 292                         buf = *tmp;
 293                 else
 294                 {
 295                         buf = create_buffer(source.size(), buffer::usage::write_only);
 296                         tbb::parallel_for(tbb::blocked_range<std::size_t>(0, source.size()), [&](const tbb::blocked_range<std::size_t>& r)
 297                         {
 298                                 A_memcpy(buf->data() + r.begin(), source.data() + r.begin(), r.size());
 299                         });
 300                 }
 301
 302                 return buf;
 303         }
 304
 305         // TODO: Since the returned texture is cached it SHOULD NOT be modified.
 306         std::future<std::shared_ptr<texture>> copy_async(const array<const std::uint8_t>& source, int width, int height, int stride, bool mipmapped)
 307         {
 308                 std::shared_ptr<buffer> buf = copy_to_buf(source);
 309
 310                 return executor_.begin_invoke([=]() -> std::shared_ptr<texture>
 311                 {
 312                         tbb::concurrent_hash_map<buffer*, std::shared_ptr<texture>>::const_accessor a;
 313                         if(texture_cache_.find(a, buf.get()))
 314                                 return spl::make_shared_ptr(a->second);
 315
 316                         auto texture = create_texture(width, height, stride, mipmapped, false);
 317                         texture->copy_from(*buf);
 318
 319                         texture_cache_.insert(std::make_pair(buf.get(), texture));
 320
 321                         return texture;
 322                 }, task_priority::high_priority);
 323         }
 324
 325         std::future<std::shared_ptr<texture>> copy_async(const array<std::uint8_t>& source, int width, int height, int stride, bool mipmapped)
 326         {
 327                 std::shared_ptr<buffer> buf = copy_to_buf(source);
 328
 329                 return executor_.begin_invoke([=]() -> std::shared_ptr<texture>
 330                 {
 331                         auto texture = create_texture(width, height, stride, mipmapped, false);
 332                         texture->copy_from(*buf);
 333
 334                         return texture;
 335                 }, task_priority::high_priority);
 336         }
 337
 338         std::future<array<const std::uint8_t>> copy_async(const spl::shared_ptr<texture>& source)
 339         {
 340                 if(!executor_.is_current())
 341                         CASPAR_THROW_EXCEPTION(invalid_operation() << msg_info("Operation only valid in an OpenGL Context."));
 342
 343                 auto buffer = create_buffer(source->size(), buffer::usage::read_only);
 344                 source->copy_to(*buffer);
 345
 346                 auto self = shared_from_this();
 347                 auto cmd = [self, buffer]() mutable -> array<const std::uint8_t>
 348                 {
 349                         self->executor_.invoke(std::bind(&buffer::map, std::ref(buffer))); // Defer blocking "map" call until data is needed.
 350                         return array<const std::uint8_t>(buffer->data(), buffer->size(), true, buffer);
 351                 };
 352                 return std::async(std::launch::deferred, std::move(cmd));
 353         }
 354
 355         std::future<void> gc()
 356         {
 357                 return executor_.begin_invoke([=]
 358                 {
 359                         CASPAR_LOG(info) << " ogl: Running GC.";
 360
 361                         try
 362                         {
 363                                 for (auto& pools : device_pools_)
 364                                 {
 365                                         for (auto& pool : pools)
 366                                                 pool.second.clear();
 367                                 }
 368                                 for (auto& pools : host_pools_)
 369                                 {
 370                                         for (auto& pool : pools)
 371                                                 pool.second.clear();
 372                                 }
 373                         }
 374                         catch (...)
 375                         {
 376                                 CASPAR_LOG_CURRENT_EXCEPTION();
 377                         }
 378                 }, task_priority::high_priority);
 379         }
 380 };
 381
 382 device::device()
 383         : executor_(L"OpenGL Rendering Context")
 384         , impl_(new impl(executor_)){}
 385 device::~device(){}
 386 spl::shared_ptr<texture>                                        device::create_texture(int width, int height, int stride, bool mipmapped){ return impl_->create_texture(width, height, stride, mipmapped, true); }
 387 array<std::uint8_t>                                                     device::create_array(int size){return impl_->create_array(size);}
 388 std::future<std::shared_ptr<texture>>           device::copy_async(const array<const std::uint8_t>& source, int width, int height, int stride, bool mipmapped){return impl_->copy_async(source, width, height, stride, mipmapped);}
 389 std::future<std::shared_ptr<texture>>           device::copy_async(const array<std::uint8_t>& source, int width, int height, int stride, bool mipmapped){ return impl_->copy_async(source, width, height, stride, mipmapped); }
 390 std::future<array<const std::uint8_t>>          device::copy_async(const spl::shared_ptr<texture>& source){return impl_->copy_async(source);}
 391 std::future<void>                                                       device::gc() { return impl_->gc(); }
 392 boost::property_tree::wptree                            device::info() const { return impl_->info(); }
 393 std::wstring                                                            device::version() const{return impl_->version();}
 394
 395
 396 }}}
 397
 398