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         void __stdcall gl_debug_callback(GLenum, GLenum, GLuint, GLenum, GLsizei, const GLchar* message, const void*) {
  60                 CASPAR_LOG(debug) << message;
  61         }
  62
  63
  64 struct device::impl : public std::enable_shared_from_this<impl>
  65 {
  66         static_assert(std::is_same<decltype(boost::declval<device>().impl_), spl::shared_ptr<impl>>::value, "impl_ must be shared_ptr");
  67
  68         tbb::concurrent_hash_map<buffer*, std::shared_ptr<texture>> texture_cache_;
  69
  70         std::unique_ptr<sf::Context> device_;
  71
  72         std::array<tbb::concurrent_unordered_map<std::size_t, tbb::concurrent_bounded_queue<std::shared_ptr<texture>>>, 8>      device_pools_;
  73         std::array<tbb::concurrent_unordered_map<std::size_t, tbb::concurrent_bounded_queue<std::shared_ptr<buffer>>>, 2>       host_pools_;
  74
  75         GLuint fbo_;
  76
  77         executor& executor_;
  78
  79         impl(executor& executor)
  80                 : executor_(executor)
  81         {
  82                 executor_.set_capacity(256);
  83
  84                 CASPAR_LOG(info) << L"Initializing OpenGL Device.";
  85
  86                 executor_.invoke([=]
  87                 {
  88                         device_.reset(new sf::Context());
  89                         device_->setActive(true);
  90
  91                         if (glewInit() != GLEW_OK)
  92                                 CASPAR_THROW_EXCEPTION(gl::ogl_exception() << msg_info("Failed to initialize GLEW."));
  93
  94                         if(!GLEW_VERSION_3_0)
  95                                 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."));
  96
  97                         glDebugMessageCallbackARB(gl_debug_callback, nullptr);
  98                         //glDebugMessageControlARB(GL_DONT_CARE, GL_DONT_CARE, GL_DEBUG_SEVERITY_NOTIFICATION, 0, nullptr, GL_FALSE);
  99                         glDebugMessageControlARB(GL_DONT_CARE, GL_DEBUG_TYPE_PERFORMANCE, GL_DEBUG_SEVERITY_LOW, 0, nullptr, GL_FALSE);
 100                         glDebugMessageControlARB(GL_DONT_CARE, GL_DEBUG_TYPE_PERFORMANCE, GL_DEBUG_SEVERITY_MEDIUM, 0, nullptr, GL_FALSE);
 101
 102                         glGenFramebuffers(1, &fbo_);
 103                         glBindFramebuffer(GL_FRAMEBUFFER, fbo_);
 104                 });
 105
 106                 CASPAR_LOG(info) << L"Successfully initialized OpenGL " << version();
 107         }
 108
 109         ~impl()
 110         {
 111                 executor_.invoke([=]
 112                 {
 113                         texture_cache_.clear();
 114
 115                         for (auto& pool : host_pools_)
 116                                 pool.clear();
 117
 118                         for (auto& pool : device_pools_)
 119                                 pool.clear();
 120
 121                         glDeleteFramebuffers(1, &fbo_);
 122
 123                         device_.reset();
 124                 });
 125         }
 126
 127         boost::property_tree::wptree info() const
 128         {
 129                 boost::property_tree::wptree info;
 130
 131                 boost::property_tree::wptree pooled_device_buffers;
 132                 size_t total_pooled_device_buffer_size  = 0;
 133                 size_t total_pooled_device_buffer_count = 0;
 134
 135                 for (size_t i = 0; i < device_pools_.size(); ++i)
 136                 {
 137                         auto& pools             = device_pools_.at(i);
 138                         bool mipmapping = i > 3;
 139                         auto stride             = mipmapping ? i - 3 : i + 1;
 140
 141                         for (auto& pool : pools)
 142                         {
 143                                 auto width      = pool.first >> 16;
 144                                 auto height     = pool.first & 0x0000FFFF;
 145                                 auto size       = width * height * stride;
 146                                 auto count      = pool.second.size();
 147
 148                                 if (count == 0)
 149                                         continue;
 150
 151                                 boost::property_tree::wptree pool_info;
 152
 153                                 pool_info.add(L"stride",                stride);
 154                                 pool_info.add(L"mipmapping",    mipmapping);
 155                                 pool_info.add(L"width",                 width);
 156                                 pool_info.add(L"height",                height);
 157                                 pool_info.add(L"size",                  size);
 158                                 pool_info.add(L"count",                 count);
 159
 160                                 total_pooled_device_buffer_size         += size * count;
 161                                 total_pooled_device_buffer_count        += count;
 162
 163                                 pooled_device_buffers.add_child(L"device_buffer_pool", pool_info);
 164                         }
 165                 }
 166
 167                 info.add_child(L"gl.details.pooled_device_buffers", pooled_device_buffers);
 168
 169                 boost::property_tree::wptree pooled_host_buffers;
 170                 size_t total_read_size          = 0;
 171                 size_t total_write_size         = 0;
 172                 size_t total_read_count         = 0;
 173                 size_t total_write_count        = 0;
 174
 175                 for (size_t i = 0; i < host_pools_.size(); ++i)
 176                 {
 177                         auto& pools     = host_pools_.at(i);
 178                         auto usage      = static_cast<buffer::usage>(i);
 179
 180                         for (auto& pool : pools)
 181                         {
 182                                 auto size       = pool.first;
 183                                 auto count      = pool.second.size();
 184
 185                                 if (count == 0)
 186                                         continue;
 187
 188                                 boost::property_tree::wptree pool_info;
 189
 190                                 pool_info.add(L"usage", usage == buffer::usage::read_only ? L"read_only" : L"write_only");
 191                                 pool_info.add(L"size",  size);
 192                                 pool_info.add(L"count", count);
 193
 194                                 pooled_host_buffers.add_child(L"host_buffer_pool", pool_info);
 195
 196                                 (usage == buffer::usage::read_only ? total_read_count : total_write_count) += count;
 197                                 (usage == buffer::usage::read_only ? total_read_size : total_write_size) += size * count;
 198                         }
 199                 }
 200
 201                 info.add_child(L"gl.details.pooled_host_buffers",                               pooled_host_buffers);
 202                 info.add(L"gl.summary.pooled_device_buffers.total_count",               total_pooled_device_buffer_count);
 203                 info.add(L"gl.summary.pooled_device_buffers.total_size",                total_pooled_device_buffer_size);
 204                 info.add_child(L"gl.summary.all_device_buffers",                                texture::info());
 205                 info.add(L"gl.summary.pooled_host_buffers.total_read_count",    total_read_count);
 206                 info.add(L"gl.summary.pooled_host_buffers.total_write_count",   total_write_count);
 207                 info.add(L"gl.summary.pooled_host_buffers.total_read_size",             total_read_size);
 208                 info.add(L"gl.summary.pooled_host_buffers.total_write_size",    total_write_size);
 209                 info.add_child(L"gl.summary.all_host_buffers",                                  buffer::info());
 210
 211                 return info;
 212         }
 213
 214         std::wstring version()
 215         {
 216                 try
 217                 {
 218                         return executor_.invoke([]
 219                         {
 220                                 return u16(reinterpret_cast<const char*>(GL2(glGetString(GL_VERSION)))) + L" " + u16(reinterpret_cast<const char*>(GL2(glGetString(GL_VENDOR))));
 221                         });
 222                 }
 223                 catch(...)
 224                 {
 225                         return L"Not found";;
 226                 }
 227         }
 228
 229         spl::shared_ptr<texture> create_texture(int width, int height, int stride, bool mipmapped, bool clear)
 230         {
 231                 CASPAR_VERIFY(stride > 0 && stride < 5);
 232                 CASPAR_VERIFY(width > 0 && height > 0);
 233
 234                 if(!executor_.is_current())
 235                         CASPAR_THROW_EXCEPTION(invalid_operation() << msg_info("Operation only valid in an OpenGL Context."));
 236
 237                 auto pool = &device_pools_[stride - 1 + (mipmapped ? 4 : 0)][((width << 16) & 0xFFFF0000) | (height & 0x0000FFFF)];
 238
 239                 std::shared_ptr<texture> tex;
 240                 if(!pool->try_pop(tex))
 241                         tex = spl::make_shared<texture>(width, height, stride, mipmapped);
 242
 243                 if(clear)
 244                         tex->clear();
 245
 246                 return spl::shared_ptr<texture>(tex.get(), [tex, pool](texture*) mutable
 247                 {
 248                         pool->push(tex);
 249                 });
 250         }
 251
 252         spl::shared_ptr<buffer> create_buffer(std::size_t size, buffer::usage usage)
 253         {
 254                 CASPAR_VERIFY(size > 0);
 255
 256                 auto pool = &host_pools_[static_cast<int>(usage)][size];
 257
 258                 std::shared_ptr<buffer> buf;
 259                 if(!pool->try_pop(buf))
 260                 {
 261                         caspar::timer timer;
 262
 263                         buf = executor_.invoke([&]
 264                         {
 265                                 return std::make_shared<buffer>(size, usage);
 266                         }, task_priority::high_priority);
 267
 268                         if(timer.elapsed() > 0.02)
 269                                 CASPAR_LOG(warning) << L"[ogl-device] Performance warning. Buffer allocation blocked: " << timer.elapsed();
 270                 }
 271
 272                 std::weak_ptr<impl> self = shared_from_this(); // buffers can leave the device context, take a hold on life-time.
 273                 return spl::shared_ptr<buffer>(buf.get(), [=](buffer*) mutable
 274                 {
 275                         auto strong = self.lock();
 276
 277                         if (strong)
 278                         {
 279                                 strong->executor_.invoke([&]
 280                                 {
 281                                         strong->texture_cache_.erase(buf.get());
 282                                 }, task_priority::high_priority);
 283
 284                                 pool->push(buf);
 285                         }
 286                         else
 287                         {
 288                                 CASPAR_LOG(info) << L"Buffer outlived ogl device";
 289                         }
 290                 });
 291         }
 292
 293         array<std::uint8_t> create_array(std::size_t size)
 294         {
 295                 auto buf = create_buffer(size, buffer::usage::write_only);
 296                 return array<std::uint8_t>(buf->data(), buf->size(), false, buf);
 297         }
 298
 299         template<typename T>
 300         std::shared_ptr<buffer> copy_to_buf(const T& source)
 301         {
 302                 std::shared_ptr<buffer> buf;
 303
 304                 auto tmp = source.template storage<spl::shared_ptr<buffer>>();
 305                 if(tmp)
 306                         buf = *tmp;
 307                 else
 308                 {
 309                         buf = create_buffer(source.size(), buffer::usage::write_only);
 310                         tbb::parallel_for(tbb::blocked_range<std::size_t>(0, source.size()), [&](const tbb::blocked_range<std::size_t>& r)
 311                         {
 312                                 A_memcpy(buf->data() + r.begin(), source.data() + r.begin(), r.size());
 313                         });
 314                 }
 315
 316                 return buf;
 317         }
 318
 319         // TODO: Since the returned texture is cached it SHOULD NOT be modified.
 320         std::future<std::shared_ptr<texture>> copy_async(const array<const std::uint8_t>& source, int width, int height, int stride, bool mipmapped)
 321         {
 322                 std::shared_ptr<buffer> buf = copy_to_buf(source);
 323
 324                 return executor_.begin_invoke([=]() -> std::shared_ptr<texture>
 325                 {
 326                         tbb::concurrent_hash_map<buffer*, std::shared_ptr<texture>>::const_accessor a;
 327                         if(texture_cache_.find(a, buf.get()))
 328                                 return spl::make_shared_ptr(a->second);
 329
 330                         auto texture = create_texture(width, height, stride, mipmapped, false);
 331                         texture->copy_from(*buf);
 332
 333                         texture_cache_.insert(std::make_pair(buf.get(), texture));
 334
 335                         return texture;
 336                 }, task_priority::high_priority);
 337         }
 338
 339         std::future<std::shared_ptr<texture>> copy_async(const array<std::uint8_t>& source, int width, int height, int stride, bool mipmapped)
 340         {
 341                 std::shared_ptr<buffer> buf = copy_to_buf(source);
 342
 343                 return executor_.begin_invoke([=]() -> std::shared_ptr<texture>
 344                 {
 345                         auto texture = create_texture(width, height, stride, mipmapped, false);
 346                         texture->copy_from(*buf);
 347
 348                         return texture;
 349                 }, task_priority::high_priority);
 350         }
 351
 352         std::future<array<const std::uint8_t>> copy_async(const spl::shared_ptr<texture>& source)
 353         {
 354                 if(!executor_.is_current())
 355                         CASPAR_THROW_EXCEPTION(invalid_operation() << msg_info("Operation only valid in an OpenGL Context."));
 356
 357                 auto buffer = create_buffer(source->size(), buffer::usage::read_only);
 358                 source->copy_to(*buffer);
 359
 360                 auto self = shared_from_this();
 361                 auto cmd = [self, buffer]() mutable -> array<const std::uint8_t>
 362                 {
 363                         self->executor_.invoke(std::bind(&buffer::map, std::ref(buffer))); // Defer blocking "map" call until data is needed.
 364                         return array<const std::uint8_t>(buffer->data(), buffer->size(), true, buffer);
 365                 };
 366                 return std::async(std::launch::deferred, std::move(cmd));
 367         }
 368
 369         std::future<void> gc()
 370         {
 371                 return executor_.begin_invoke([=]
 372                 {
 373                         CASPAR_LOG(info) << " ogl: Running GC.";
 374
 375                         try
 376                         {
 377                                 for (auto& pools : device_pools_)
 378                                 {
 379                                         for (auto& pool : pools)
 380                                                 pool.second.clear();
 381                                 }
 382                                 for (auto& pools : host_pools_)
 383                                 {
 384                                         for (auto& pool : pools)
 385                                                 pool.second.clear();
 386                                 }
 387                         }
 388                         catch (...)
 389                         {
 390                                 CASPAR_LOG_CURRENT_EXCEPTION();
 391                         }
 392                 }, task_priority::high_priority);
 393         }
 394 };
 395
 396 device::device()
 397         : executor_(L"OpenGL Rendering Context")
 398         , impl_(new impl(executor_)){}
 399 device::~device(){}
 400 spl::shared_ptr<texture>                                        device::create_texture(int width, int height, int stride, bool mipmapped){ return impl_->create_texture(width, height, stride, mipmapped, true); }
 401 array<std::uint8_t>                                                     device::create_array(int size){return impl_->create_array(size);}
 402 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);}
 403 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); }
 404 std::future<array<const std::uint8_t>>          device::copy_async(const spl::shared_ptr<texture>& source){return impl_->copy_async(source);}
 405 std::future<void>                                                       device::gc() { return impl_->gc(); }
 406 boost::property_tree::wptree                            device::info() const { return impl_->info(); }
 407 std::wstring                                                            device::version() const{return impl_->version();}
 408
 409
 410 }}}
 411
 412