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

   1 /*\r
   2 * Copyright (c) 2011 Sveriges Television AB <info@casparcg.com>\r
   3 *\r
   4 * This file is part of CasparCG (www.casparcg.com).\r
   5 *\r
   6 * CasparCG is free software: you can redistribute it and/or modify\r
   7 * it under the terms of the GNU General Public License as published by\r
   8 * the Free Software Foundation, either version 3 of the License, or\r
   9 * (at your option) any later version.\r
  10 *\r
  11 * CasparCG is distributed in the hope that it will be useful,\r
  12 * but WITHOUT ANY WARRANTY; without even the implied warranty of\r
  13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the\r
  14 * GNU General Public License for more details.\r
  15 *\r
  16 * You should have received a copy of the GNU General Public License\r
  17 * along with CasparCG. If not, see <http://www.gnu.org/licenses/>.\r
  18 *\r
  19 * Author: Robert Nagy, ronag89@gmail.com\r
  20 */\r
  21 \r
  22 // TODO: Smart GC\r
  23 \r
  24 #include "../../stdafx.h"\r
  25 \r
  26 #include "device.h"\r
  27 \r
  28 #include "buffer.h"\r
  29 #include "texture.h"\r
  30 #include "shader.h"\r
  31 \r
  32 #include <common/assert.h>\r
  33 #include <common/except.h>\r
  34 #include <common/concurrency/async.h>\r
  35 #include <common/memory/array.h>\r
  36 #include <common/gl/gl_check.h>\r
  37 #include <common/os/windows/windows.h>\r
  38 \r
  39 #include <boost/foreach.hpp>\r
  40 \r
  41 #include <gl/glew.h>\r
  42 \r
  43 #include <SFML/Window/Context.hpp>\r
  44 \r
  45 #include <tbb/concurrent_unordered_map.h>\r
  46 #include <tbb/concurrent_hash_map.h>\r
  47 #include <tbb/concurrent_queue.h>\r
  48 \r
  49 #include <boost/utility/declval.hpp>\r
  50 \r
  51 #include <array>\r
  52 #include <unordered_map>\r
  53 \r
  54 tbb::atomic<int> g_count = tbb::atomic<int>();\r
  55 \r
  56 namespace caspar { namespace accelerator { namespace ogl {\r
  57                 \r
  58 struct device::impl : public std::enable_shared_from_this<impl>\r
  59 {       \r
  60         static_assert(std::is_same<decltype(boost::declval<device>().impl_), spl::shared_ptr<impl>>::value, "impl_ must be shared_ptr");\r
  61 \r
  62         tbb::concurrent_hash_map<buffer*, std::shared_ptr<texture>> texture_cache_;\r
  63 \r
  64         std::unique_ptr<sf::Context> device_;\r
  65         std::unique_ptr<sf::Context> host_alloc_device_;\r
  66         \r
  67         std::array<tbb::concurrent_unordered_map<std::size_t, tbb::concurrent_bounded_queue<std::shared_ptr<texture>>>, 4>      device_pools_;\r
  68         std::array<tbb::concurrent_unordered_map<std::size_t, tbb::concurrent_bounded_queue<std::shared_ptr<buffer>>>, 2>       host_pools_;\r
  69         \r
  70         GLuint fbo_;\r
  71 \r
  72         executor& render_executor_;\r
  73         executor  alloc_executor_;\r
  74                                 \r
  75         impl(executor& executor) \r
  76                 : render_executor_(executor)\r
  77                 , alloc_executor_(L"OpenGL allocation context.")\r
  78         {\r
  79                 if(g_count++ > 1)\r
  80                         CASPAR_LOG(warning) << L"Multiple OGL devices.";\r
  81 \r
  82                 CASPAR_LOG(info) << L"Initializing OpenGL Device.";\r
  83                 \r
  84                 auto ctx1 = render_executor_.invoke([=]() -> HGLRC \r
  85                 {\r
  86                         device_.reset(new sf::Context());\r
  87                         device_->SetActive(true);               \r
  88                                                 \r
  89                         if (glewInit() != GLEW_OK)\r
  90                                 BOOST_THROW_EXCEPTION(gl::ogl_exception() << msg_info("Failed to initialize GLEW."));\r
  91                 \r
  92                         if(!GLEW_VERSION_3_0)\r
  93                                 BOOST_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."));\r
  94         \r
  95                         glGenFramebuffers(1, &fbo_);                            \r
  96                         glBindFramebuffer(GL_FRAMEBUFFER, fbo_);\r
  97                         \r
  98                         auto ctx1 = wglGetCurrentContext();\r
  99                         \r
 100                         device_->SetActive(false);\r
 101 \r
 102                         return ctx1;\r
 103                 });\r
 104 \r
 105                 alloc_executor_.invoke([=]\r
 106                 {\r
 107                         host_alloc_device_.reset(new sf::Context());\r
 108                         host_alloc_device_->SetActive(true);    \r
 109                         auto ctx2 = wglGetCurrentContext();\r
 110 \r
 111                         if(!wglShareLists(ctx1, ctx2))\r
 112                                 BOOST_THROW_EXCEPTION(gl::ogl_exception() << msg_info("Failed to share OpenGL devices."));\r
 113                 });\r
 114 \r
 115                 render_executor_.invoke([=]\r
 116                 {               \r
 117                         device_->SetActive(true);\r
 118                 });\r
 119                 \r
 120                 CASPAR_LOG(info) << L"Successfully initialized OpenGL " << version();\r
 121         }\r
 122 \r
 123         ~impl()\r
 124         {\r
 125                 alloc_executor_.invoke([=]\r
 126                 {\r
 127                         host_alloc_device_.reset();\r
 128                         BOOST_FOREACH(auto& pool, host_pools_)\r
 129                                 pool.clear();\r
 130                 });\r
 131 \r
 132                 render_executor_.invoke([=]\r
 133                 {\r
 134                         BOOST_FOREACH(auto& pool, device_pools_)\r
 135                                 pool.clear();\r
 136                         glDeleteFramebuffers(1, &fbo_);\r
 137 \r
 138                         device_.reset();\r
 139                 });\r
 140         }\r
 141                 \r
 142         std::wstring version()\r
 143         {       \r
 144                 try\r
 145                 {\r
 146                         return alloc_executor_.invoke([]\r
 147                         {\r
 148                                 return u16(reinterpret_cast<const char*>(GL2(glGetString(GL_VERSION)))) + L" " + u16(reinterpret_cast<const char*>(GL2(glGetString(GL_VENDOR))));\r
 149                         });     \r
 150                 }\r
 151                 catch(...)\r
 152                 {\r
 153                         return L"Not found";;\r
 154                 }\r
 155         }\r
 156                                                         \r
 157         spl::shared_ptr<texture> create_texture(int width, int height, int stride)\r
 158         {\r
 159                 CASPAR_VERIFY(stride > 0 && stride < 5);\r
 160                 CASPAR_VERIFY(width > 0 && height > 0);\r
 161                 \r
 162                 auto pool = &device_pools_[stride-1][((width << 16) & 0xFFFF0000) | (height & 0x0000FFFF)];\r
 163                 \r
 164                 std::shared_ptr<texture> buffer;\r
 165                 if(!pool->try_pop(buffer))              \r
 166                         buffer = spl::make_shared<texture>(width, height, stride);\r
 167         \r
 168                 return spl::shared_ptr<texture>(buffer.get(), [buffer, pool](texture*) mutable\r
 169                 {               \r
 170                         pool->push(buffer);     \r
 171                 });\r
 172         }\r
 173                 \r
 174         spl::shared_ptr<buffer> create_buffer(std::size_t size, buffer::usage usage)\r
 175         {\r
 176                 CASPAR_VERIFY(size > 0);\r
 177                 \r
 178                 auto pool = &host_pools_[usage.value()][size];\r
 179                 \r
 180                 std::shared_ptr<buffer> buf;\r
 181                 if(!pool->try_pop(buf)) \r
 182                 {\r
 183                         buf = alloc_executor_.invoke([&]\r
 184                         {\r
 185                                 return spl::make_shared<buffer>(size, usage);\r
 186                         });\r
 187                 }\r
 188                                 \r
 189                 auto ptr = buf->data();\r
 190                 auto self = shared_from_this(); // buffers can leave the device context, take a hold on life-time.\r
 191 \r
 192                 auto on_release = [self, buf, ptr, usage, pool]() mutable\r
 193                 {               \r
 194                         if(usage == buffer::usage::write_only)                                  \r
 195                                 buf->map();                                     \r
 196                         else\r
 197                                 buf->unmap();\r
 198 \r
 199                         self->texture_cache_.erase(buf.get());\r
 200 \r
 201                         pool->push(buf);\r
 202                 };\r
 203                 \r
 204                 return spl::shared_ptr<buffer>(buf.get(), [=](buffer*) mutable\r
 205                 {\r
 206                         self->alloc_executor_.begin_invoke(on_release); \r
 207                 });\r
 208         }\r
 209 \r
 210         array<std::uint8_t> create_array(std::size_t size)\r
 211         {               \r
 212                 auto buf = create_buffer(size, buffer::usage::write_only);\r
 213                 return array<std::uint8_t>(buf->data(), buf->size(), false, buf);\r
 214         }\r
 215 \r
 216         boost::unique_future<spl::shared_ptr<texture>> copy_async(const array<const std::uint8_t>& source, int width, int height, int stride)\r
 217         {\r
 218                 auto buf = source.storage<spl::shared_ptr<buffer>>();\r
 219                                 \r
 220                 return render_executor_.begin_invoke([=]() -> spl::shared_ptr<texture>\r
 221                 {\r
 222                         tbb::concurrent_hash_map<buffer*, std::shared_ptr<texture>>::const_accessor a;\r
 223                         if(texture_cache_.find(a, buf.get()))\r
 224                                 return spl::make_shared_ptr(a->second);\r
 225 \r
 226                         auto texture = create_texture(width, height, stride);\r
 227                         texture->copy_from(*buf);       \r
 228 \r
 229                         texture_cache_.insert(std::make_pair(buf.get(), texture));\r
 230 \r
 231                         return texture;\r
 232 \r
 233                 }, task_priority::high_priority);\r
 234         }\r
 235         \r
 236         boost::unique_future<spl::shared_ptr<texture>> copy_async(const array<std::uint8_t>& source, int width, int height, int stride)\r
 237         {\r
 238                 auto buf = source.storage<spl::shared_ptr<buffer>>();\r
 239                                 \r
 240                 return render_executor_.begin_invoke([=]() -> spl::shared_ptr<texture>\r
 241                 {\r
 242                         auto texture = create_texture(width, height, stride);\r
 243                         texture->copy_from(*buf);                               \r
 244                         return texture;\r
 245                 }, task_priority::high_priority);\r
 246         }\r
 247 \r
 248         boost::unique_future<array<const std::uint8_t>> copy_async(const spl::shared_ptr<texture>& source)\r
 249         {\r
 250                 return flatten(render_executor_.begin_invoke([=]() -> boost::shared_future<array<const std::uint8_t>>\r
 251                 {\r
 252                         auto buffer = create_buffer(source->size(), buffer::usage::read_only); \r
 253                         source->copy_to(*buffer);       \r
 254 \r
 255                         auto self = shared_from_this();\r
 256                         return make_shared(async(launch::deferred, [self, buffer]() mutable -> array<const std::uint8_t>\r
 257                         {\r
 258                                 const auto& buf = buffer.get();\r
 259                                 if(!buf->data())\r
 260                                         self->alloc_executor_.invoke(std::bind(&buffer::map, std::ref(buf))); // Defer blocking "map" call until data is needed.\r
 261 \r
 262                                 return array<const std::uint8_t>(buf->data(), buf->size(), true, buffer);\r
 263                         }));\r
 264                 }, task_priority::high_priority));\r
 265         }\r
 266 };\r
 267 \r
 268 device::device() \r
 269         : executor_(L"OpenGL Rendering Context.")\r
 270         , impl_(new impl(executor_)){}\r
 271 device::~device(){}\r
 272 spl::shared_ptr<texture>                                                        device::create_texture(int width, int height, int stride){return impl_->create_texture(width, height, stride);}\r
 273 array<std::uint8_t>                                                                     device::create_array(int size){return impl_->create_array(size);}\r
 274 boost::unique_future<spl::shared_ptr<texture>>          device::copy_async(const array<const std::uint8_t>& source, int width, int height, int stride){return impl_->copy_async(source, width, height, stride);}\r
 275 boost::unique_future<spl::shared_ptr<texture>>          device::copy_async(const array<std::uint8_t>& source, int width, int height, int stride){return impl_->copy_async(source, width, height, stride);}\r
 276 boost::unique_future<array<const std::uint8_t>>         device::copy_async(const spl::shared_ptr<texture>& source){return impl_->copy_async(source);}\r
 277 std::wstring                                                                            device::version() const{return impl_->version();}\r
 278 \r
 279 \r
 280 }}}\r
 281 \r
 282 \r