| 1 | // Copyright (C) 2016 The Qt Company Ltd. |
| 2 | // SPDX-License-Identifier: LicenseRef-Qt-Commercial OR LGPL-3.0-only OR GPL-2.0-only OR GPL-3.0-only |
| 3 | |
| 4 | #include "qsgrendernode.h" |
| 5 | #include "qsgrendernode_p.h" |
| 6 | |
| 7 | QT_BEGIN_NAMESPACE |
| 8 | |
| 9 | /*! |
| 10 | \class QSGRenderNode |
| 11 | \brief The QSGRenderNode class represents a set of custom rendering commands |
| 12 | targeting the graphics API that is in use by the scenegraph. |
| 13 | \inmodule QtQuick |
| 14 | \since 5.8 |
| 15 | |
| 16 | QSGRenderNode allows creating scene graph nodes that perform their own |
| 17 | custom rendering via QRhi (the common approach from Qt 6.6 on), directly |
| 18 | via a 3D graphics API such as OpenGL, Vulkan, or Metal, or, when the \c |
| 19 | software backend is in use, via QPainter. |
| 20 | |
| 21 | QSGRenderNode is the enabler for one of the three ways to integrate custom |
| 22 | 2D/3D rendering into a Qt Quick scene. The other two options are to perform |
| 23 | the rendering \c before or \c after the Qt Quick scene's own rendering, |
| 24 | or to generate a whole separate render pass targeting a dedicated render |
| 25 | target (a texture) and then have an item in the scene display the texture. |
| 26 | The QSGRenderNode-based approach is similar to the former, in the sense |
| 27 | that no additional render passes or render targets are involved, and allows |
| 28 | injecting custom rendering commands "inline" with the Qt Quick scene's |
| 29 | own rendering. |
| 30 | |
| 31 | \sa {Scene Graph - Custom QSGRenderNode} |
| 32 | */ |
| 33 | |
| 34 | QSGRenderNode::QSGRenderNode() |
| 35 | : QSGNode(RenderNodeType), |
| 36 | d(new QSGRenderNodePrivate) |
| 37 | { |
| 38 | } |
| 39 | |
| 40 | /*! |
| 41 | Destructs the render node. Derived classes are expected to perform cleanup |
| 42 | similar to releaseResources() in here. |
| 43 | |
| 44 | When a low-level graphics API is in use, the scenegraph will make sure |
| 45 | there is a CPU-side wait for the GPU to complete all work submitted to the |
| 46 | scenegraph's graphics command queue before the scenegraph's nodes are |
| 47 | deleted. Therefore there is no need to issue additional waits here, unless |
| 48 | the render() implementation is using additional command queues. |
| 49 | |
| 50 | With QRhi and resources such as QRhiBuffer, QRhiTexture, |
| 51 | QRhiGraphicsPipeline, etc., it is often good practice to use smart |
| 52 | pointers, such as std::unique_ptr, which can often avoid the need to |
| 53 | implement a destructor, and lead to more compact source code. Keep in mind |
| 54 | however that implementing releaseResources(), most likely issuing a number |
| 55 | of reset() calls on the unique_ptrs, is still important. |
| 56 | |
| 57 | \sa releaseResources() |
| 58 | */ |
| 59 | QSGRenderNode::~QSGRenderNode() |
| 60 | { |
| 61 | delete d; |
| 62 | } |
| 63 | |
| 64 | QSGRenderNodePrivate::QSGRenderNodePrivate() |
| 65 | : m_matrix(nullptr) |
| 66 | , m_clip_list(nullptr) |
| 67 | , m_opacity(1) |
| 68 | { |
| 69 | m_projectionMatrix.resize(sz: 1); |
| 70 | } |
| 71 | |
| 72 | /*! |
| 73 | When the underlying rendering API is OpenGL, this function should return a |
| 74 | mask where each bit represents graphics states changed by the \l render() |
| 75 | function: |
| 76 | |
| 77 | \value DepthState depth write mask, depth test enabled, depth comparison function |
| 78 | \value StencilState stencil write masks, stencil test enabled, stencil operations, |
| 79 | stencil comparison functions |
| 80 | \value ScissorState scissor enabled, scissor test enabled |
| 81 | \value ColorState clear color, color write mask |
| 82 | \value BlendState blend enabled, blend function |
| 83 | \value CullState front face, cull face enabled |
| 84 | \value ViewportState viewport |
| 85 | \value RenderTargetState render target |
| 86 | |
| 87 | With APIs other than OpenGL, the only relevant values are the ones that |
| 88 | correspond to dynamic state changes recorded on the command list/buffer. |
| 89 | For example, RSSetViewports, RSSetScissorRects, OMSetBlendState, |
| 90 | OMSetDepthStencilState in case of D3D11, or vkCmdSetViewport, vkCmdSetScissor, |
| 91 | vkCmdSetBlendConstants, vkCmdSetStencilRef in case of Vulkan, and only when |
| 92 | such commands were added to the scenegraph's command list queried via the |
| 93 | QSGRendererInterface::CommandList resource enum. States set in pipeline |
| 94 | state objects do not need to be reported here. Similarly, draw call related |
| 95 | settings (pipeline states, descriptor sets, vertex or index buffer |
| 96 | bindings, root signature, descriptor heaps, etc.) are always set again by |
| 97 | the scenegraph so render() can freely change them. |
| 98 | |
| 99 | RenderTargetState is no longer supported with APIs like Vulkan. This |
| 100 | is by nature. render() is invoked while the Qt Quick scenegraph's main |
| 101 | command buffer is recording a renderpass, so there is no possibility of |
| 102 | changing the target and starting another renderpass (on that command buffer |
| 103 | at least). Therefore returning a value with RenderTargetState set is not |
| 104 | sensible. |
| 105 | |
| 106 | \note The \c software backend exposes its QPainter and saves and restores |
| 107 | before and after invoking render(). Therefore reporting any changed states |
| 108 | from here is not necessary. |
| 109 | |
| 110 | The function is called by the renderer so it can reset the states after |
| 111 | rendering this node. This makes the implementation of render() simpler |
| 112 | since it does not have to query and restore these states. |
| 113 | |
| 114 | The default implementation returns 0, meaning no relevant state was changed |
| 115 | in render(). |
| 116 | |
| 117 | \note This function may be called before render(). |
| 118 | |
| 119 | \note With Qt 6 and QRhi-based rendering the only relevant values are |
| 120 | ViewportState and ScissorState. Other values can be returned but are |
| 121 | ignored in practice. |
| 122 | */ |
| 123 | QSGRenderNode::StateFlags QSGRenderNode::changedStates() const |
| 124 | { |
| 125 | return {}; |
| 126 | } |
| 127 | |
| 128 | /*! |
| 129 | Called from the frame preparation phase. There is a call to this function |
| 130 | before each invocation of render(). |
| 131 | |
| 132 | Unlike render(), this function is called before the scenegraph starts |
| 133 | recording the render pass for the current frame on the underlying command |
| 134 | buffer. This is useful when doing rendering with graphics APIs, such as |
| 135 | Vulkan, where copy type of operations will need to be recorded before the |
| 136 | render pass. |
| 137 | |
| 138 | The default implementation is empty. |
| 139 | |
| 140 | When implementing a QSGRenderNode that uses QRhi to render, query the QRhi |
| 141 | object from the QQuickWindow via \l{QQuickWindow::rhi()}. To get a |
| 142 | QRhiCommandBuffer for submitting work to, call commandBuffer(). To query |
| 143 | information about the active render target, call renderTarget(). See the |
| 144 | \l{{Scene Graph - Custom QSGRenderNode}} example for details. |
| 145 | |
| 146 | \since 6.0 |
| 147 | */ |
| 148 | void QSGRenderNode::prepare() |
| 149 | { |
| 150 | } |
| 151 | |
| 152 | /*! |
| 153 | \fn void QSGRenderNode::render(const RenderState *state) |
| 154 | |
| 155 | This function is called by the renderer and should paint this node with |
| 156 | directly invoking commands in the graphics API (OpenGL, Direct3D, etc.) |
| 157 | currently in use. |
| 158 | |
| 159 | The effective opacity can be retrieved with \l inheritedOpacity(). |
| 160 | |
| 161 | The projection matrix is available through \a state, while the model-view |
| 162 | matrix can be fetched with \l matrix(). The combined matrix is then the |
| 163 | projection matrix times the model-view matrix. The correct stacking of the |
| 164 | items in the scene is ensured by the projection matrix. |
| 165 | |
| 166 | When using the provided matrices, the coordinate system for vertex data |
| 167 | follows the usual QQuickItem conventions: top-left is (0, 0), bottom-right |
| 168 | is the corresponding QQuickItem's width() and height() minus one. For |
| 169 | example, assuming a two float (x-y) per vertex coordinate layout, a |
| 170 | triangle covering half of the item can be specified as (width - 1, height - 1), |
| 171 | (0, 0), (0, height - 1) using counter-clockwise direction. |
| 172 | |
| 173 | \note QSGRenderNode is provided as a means to implement custom 2D or 2.5D |
| 174 | Qt Quick items. It is not intended for integrating true 3D content into the |
| 175 | Qt Quick scene. That use case is better supported by |
| 176 | QQuickFramebufferObject, QQuickWindow::beforeRendering(), or the |
| 177 | equivalents of those for APIs other than OpenGL. |
| 178 | |
| 179 | \note QSGRenderNode can perform significantly better than texture-based |
| 180 | approaches (such as, QQuickFramebufferObject), especially on systems where |
| 181 | the fragment processing power is limited. This is because it avoids |
| 182 | rendering to a texture and then drawing a textured quad. Rather, |
| 183 | QSGRenderNode allows recording draw calls in line with the scenegraph's |
| 184 | other commands, avoiding an additional render target and the potentially |
| 185 | expensive texturing and blending. |
| 186 | |
| 187 | Clip information is calculated before the function is called. |
| 188 | Implementations wishing to take clipping into account can set up scissoring |
| 189 | or stencil based on the information in \a state. The stencil buffer is |
| 190 | filled with the necessary clip shapes, but it is up to the implementation |
| 191 | to enable stencil testing. |
| 192 | |
| 193 | Some scenegraph backends, software in particular, use no scissor or |
| 194 | stencil. There the clip region is provided as an ordinary QRegion. |
| 195 | |
| 196 | When implementing a QSGRenderNode that uses QRhi to render, query the QRhi |
| 197 | object from the QQuickWindow via \l{QQuickWindow::rhi()}. To get a |
| 198 | QRhiCommandBuffer for submitting work to, call commandBuffer(). To query |
| 199 | information about the active render target, call renderTarget(). See the |
| 200 | \l{{Scene Graph - Custom QSGRenderNode}} example for details. |
| 201 | |
| 202 | With Qt 6 and its QRhi-based scene graph renderer, no assumptions should be |
| 203 | made about the active (OpenGL) state when this function is called, even |
| 204 | when OpenGL is in use. Assume nothing about the pipelines and dynamic |
| 205 | states bound on the command list/buffer when this function is called. |
| 206 | |
| 207 | \note Depth writes are expected to be disabled. Enabling depth writes can |
| 208 | lead to unexpected results, depending on the scenegraph backend in use and |
| 209 | the content in the scene, so exercise caution with this. |
| 210 | |
| 211 | \note In Qt 6, \l changedStates() has limited use. See the documentation |
| 212 | for changedStates() for more information. |
| 213 | |
| 214 | With some graphics APIs, including when using QRhi directly, it can be |
| 215 | necessary to reimplement prepare() in addition, or alternatively connect to |
| 216 | the QQuickWindow::beforeRendering() signal. These are called/emitted before |
| 217 | recording the beginning of a renderpass on the command buffer |
| 218 | (vkCmdBeginRenderPass with Vulkan, or starting to encode via |
| 219 | MTLRenderCommandEncoder in case of Metal. Recording copy operations cannot |
| 220 | be done inside render() with such APIs. Rather, do such operations either |
| 221 | in prepare() or the slot connected to beforeRendering (with |
| 222 | DirectConnection). |
| 223 | |
| 224 | \sa QSGRendererInterface, QQuickWindow::rendererInterface() |
| 225 | */ |
| 226 | |
| 227 | /*! |
| 228 | This function is called when all custom graphics resources allocated by |
| 229 | this node have to be freed immediately. In case the node does not directly |
| 230 | allocate graphics resources (buffers, textures, render targets, fences, |
| 231 | etc.) through the graphics API that is in use, there is nothing to do here. |
| 232 | |
| 233 | Failing to release all custom resources can lead to incorrect behavior in |
| 234 | graphics device loss scenarios on some systems since subsequent |
| 235 | reinitialization of the graphics system may fail. |
| 236 | |
| 237 | \note Some scenegraph backends may choose not to call this function. |
| 238 | Therefore it is expected that QSGRenderNode implementations perform cleanup |
| 239 | both in their destructor and in releaseResources(). |
| 240 | |
| 241 | Unlike with the destructor, it is expected that render() can reinitialize |
| 242 | all resources it needs when called after a call to releaseResources(). |
| 243 | |
| 244 | With OpenGL, the scenegraph's OpenGL context will be current both when |
| 245 | calling the destructor and this function. |
| 246 | */ |
| 247 | void QSGRenderNode::releaseResources() |
| 248 | { |
| 249 | } |
| 250 | |
| 251 | /*! |
| 252 | \enum QSGRenderNode::StateFlag |
| 253 | |
| 254 | This enum is a bit mask identifying several states. |
| 255 | |
| 256 | \value DepthState Depth |
| 257 | \value StencilState Stencil |
| 258 | \value ScissorState Scissor |
| 259 | \value ColorState Color |
| 260 | \value BlendState Blend |
| 261 | \value CullState Cull |
| 262 | \value ViewportState View poirt |
| 263 | \value RenderTargetState Render target |
| 264 | |
| 265 | */ |
| 266 | |
| 267 | /*! |
| 268 | \enum QSGRenderNode::RenderingFlag |
| 269 | |
| 270 | Possible values for the bitmask returned from flags(). |
| 271 | |
| 272 | \value BoundedRectRendering Indicates that the implementation of render() |
| 273 | does not render outside the area reported from rect() in item |
| 274 | coordinates. Such node implementations can lead to more efficient rendering, |
| 275 | depending on the scenegraph backend. For example, the \c software backend can |
| 276 | continue to use the more optimal partial update path when all render nodes |
| 277 | in the scene have this flag set. |
| 278 | |
| 279 | \value DepthAwareRendering Indicates that the implementations of render() |
| 280 | conforms to scenegraph expectations by only generating a Z value of 0 in |
| 281 | scene coordinates which is then transformed by the matrices retrieved from |
| 282 | RenderState::projectionMatrix() and matrix(), as described in the notes for |
| 283 | render(). Such node implementations can lead to more efficient rendering, |
| 284 | depending on the scenegraph backend. For example, the batching OpenGL |
| 285 | renderer can continue to use a more optimal path when all render nodes in |
| 286 | the scene have this flag set. |
| 287 | |
| 288 | \value OpaqueRendering Indicates that the implementation of render() writes |
| 289 | out opaque pixels for the entire area reported from rect(). By default the |
| 290 | renderers must assume that render() can also output semi or fully |
| 291 | transparent pixels. Setting this flag can improve performance in some |
| 292 | cases. |
| 293 | |
| 294 | \value NoExternalRendering Indicates that the implementation of prepare() |
| 295 | and render() use the QRhi family of APIs, instead of directly calling a 3D |
| 296 | API such as OpenGL, Vulkan, or Metal. |
| 297 | |
| 298 | \sa render(), prepare(), rect(), QRhi |
| 299 | */ |
| 300 | |
| 301 | /*! |
| 302 | \return flags describing the behavior of this render node. |
| 303 | |
| 304 | The default implementation returns 0. |
| 305 | |
| 306 | \sa RenderingFlag, rect() |
| 307 | */ |
| 308 | QSGRenderNode::RenderingFlags QSGRenderNode::flags() const |
| 309 | { |
| 310 | return {}; |
| 311 | } |
| 312 | |
| 313 | /*! |
| 314 | \return the bounding rectangle in item coordinates for the area render() |
| 315 | touches. The value is only in use when flags() includes |
| 316 | BoundedRectRendering, ignored otherwise. |
| 317 | |
| 318 | Reporting the rectangle in combination with BoundedRectRendering is |
| 319 | particularly important with the \c software backend because otherwise |
| 320 | having a rendernode in the scene would trigger fullscreen updates, skipping |
| 321 | all partial update optimizations. |
| 322 | |
| 323 | For rendernodes covering the entire area of a corresponding QQuickItem the |
| 324 | return value will be (0, 0, item->width(), item->height()). |
| 325 | |
| 326 | \note Nodes are also free to render outside the boundaries specified by the |
| 327 | item's width and height, since the scenegraph nodes are not bounded by the |
| 328 | QQuickItem geometry, as long as this is reported correctly from this function. |
| 329 | |
| 330 | \sa flags() |
| 331 | */ |
| 332 | QRectF QSGRenderNode::rect() const |
| 333 | { |
| 334 | return QRectF(); |
| 335 | } |
| 336 | |
| 337 | /*! |
| 338 | \return pointer to the current projection matrix. |
| 339 | |
| 340 | In render() this is the same matrix that is returned from |
| 341 | RenderState::projectionMatrix(). This getter exists so that prepare() also |
| 342 | has a way to query the projection matrix. |
| 343 | |
| 344 | When working with a modern graphics API, or Qt's own graphics abstraction |
| 345 | layer, it is more than likely that one will want to load |
| 346 | \c{*projectionMatrix() * *matrix()} into a uniform buffer. That is however |
| 347 | something that needs to be done in prepare(), so outside the recording of a |
| 348 | render pass. That is why both matrices are queriable directly from the |
| 349 | QSGRenderNode, both in prepare() and render(). |
| 350 | |
| 351 | \since 6.5 |
| 352 | */ |
| 353 | const QMatrix4x4 *QSGRenderNode::projectionMatrix() const |
| 354 | { |
| 355 | return &d->m_projectionMatrix[0]; |
| 356 | } |
| 357 | |
| 358 | /*! |
| 359 | \internal |
| 360 | */ |
| 361 | const QMatrix4x4 *QSGRenderNode::projectionMatrix(qsizetype index) const |
| 362 | { |
| 363 | return &d->m_projectionMatrix[index]; |
| 364 | } |
| 365 | |
| 366 | /*! |
| 367 | \return pointer to the current model-view matrix. |
| 368 | */ |
| 369 | const QMatrix4x4 *QSGRenderNode::matrix() const |
| 370 | { |
| 371 | return d->m_matrix; |
| 372 | } |
| 373 | |
| 374 | /*! |
| 375 | \return the current clip list. |
| 376 | */ |
| 377 | const QSGClipNode *QSGRenderNode::clipList() const |
| 378 | { |
| 379 | return d->m_clip_list; |
| 380 | } |
| 381 | |
| 382 | /*! |
| 383 | \return the current effective opacity. |
| 384 | */ |
| 385 | qreal QSGRenderNode::inheritedOpacity() const |
| 386 | { |
| 387 | return d->m_opacity; |
| 388 | } |
| 389 | |
| 390 | /*! |
| 391 | \return the current render target. |
| 392 | |
| 393 | This is provided mainly to enable prepare() and render() implementations |
| 394 | that use QRhi accessing the QRhiRenderTarget's |
| 395 | \l{QRhiRenderPassDescriptor}{renderPassDescriptor} or |
| 396 | \l{QRhiRenderTarget::pixelSize()}{pixel size}. |
| 397 | |
| 398 | To build a QRhiGraphicsPipeline, which implies having to provide a |
| 399 | QRhiRenderPassDescriptor, query the renderPassDescriptor from the render |
| 400 | target. Be aware however that the render target may change over the |
| 401 | lifetime of the custom QQuickItem and the QSGRenderNode. For example, |
| 402 | consider what happens when dynamically setting \c{layer.enabled: true} on |
| 403 | the item or an ancestor of it: this triggers rendering into a texture, not |
| 404 | directly to the window, which means the QSGRenderNode is going to work with |
| 405 | a different render target from then on. The new render target may then have |
| 406 | a different pixel format, which can make already built graphics pipelines |
| 407 | incompatible. This can be handled with logic such as the following: |
| 408 | |
| 409 | \code |
| 410 | if (m_pipeline && renderTarget()->renderPassDescriptor()->serializedFormat() != m_renderPassFormat) { |
| 411 | delete m_pipeline; |
| 412 | m_pipeline = nullptr; |
| 413 | } |
| 414 | if (!m_pipeline) { |
| 415 | // Build a new QRhiGraphicsPipeline. |
| 416 | // ... |
| 417 | // Store the serialized format for fast and simple comparisons later on. |
| 418 | m_renderPassFormat = renderTarget()->renderPassDescriptor()->serializedFormat(); |
| 419 | } |
| 420 | \endcode |
| 421 | |
| 422 | \since 6.6 |
| 423 | |
| 424 | \sa commandBuffer() |
| 425 | */ |
| 426 | QRhiRenderTarget *QSGRenderNode::renderTarget() const |
| 427 | { |
| 428 | return d->m_rt.rt; |
| 429 | } |
| 430 | |
| 431 | /*! |
| 432 | \return the current command buffer. |
| 433 | |
| 434 | \since 6.6 |
| 435 | |
| 436 | \sa renderTarget() |
| 437 | */ |
| 438 | QRhiCommandBuffer *QSGRenderNode::commandBuffer() const |
| 439 | { |
| 440 | return d->m_rt.cb; |
| 441 | } |
| 442 | |
| 443 | QSGRenderNode::RenderState::~RenderState() |
| 444 | { |
| 445 | } |
| 446 | |
| 447 | /*! |
| 448 | \fn const QMatrix4x4 *QSGRenderNode::RenderState::projectionMatrix() const |
| 449 | |
| 450 | \return pointer to the current projection matrix. |
| 451 | |
| 452 | The model-view matrix can be retrieved with QSGRenderNode::matrix(). |
| 453 | Typically \c{projection * modelview} is the matrix that is then used in the |
| 454 | vertex shader to transform the vertices. |
| 455 | */ |
| 456 | |
| 457 | /*! |
| 458 | \fn const QMatrix4x4 *QSGRenderNode::RenderState::scissorRect() const |
| 459 | |
| 460 | \return the current scissor rectangle when clipping is active. x and y are |
| 461 | the bottom left coordinates. |
| 462 | */ |
| 463 | |
| 464 | /*! |
| 465 | \fn const QMatrix4x4 *QSGRenderNode::RenderState::scissorEnabled() const |
| 466 | |
| 467 | \return the current state of scissoring. |
| 468 | |
| 469 | \note Only relevant for graphics APIs that have a dedicated on/off state of |
| 470 | scissoring. |
| 471 | */ |
| 472 | |
| 473 | /*! |
| 474 | \fn const QMatrix4x4 *QSGRenderNode::RenderState::stencilValue() const |
| 475 | |
| 476 | \return the current stencil reference value when clipping is active. |
| 477 | */ |
| 478 | |
| 479 | /*! |
| 480 | \fn const QMatrix4x4 *QSGRenderNode::RenderState::stencilEnabled() const |
| 481 | |
| 482 | \return the current state of stencil testing. |
| 483 | |
| 484 | \note With graphics APIs where stencil testing is enabled in pipeline state |
| 485 | objects, instead of individual state-setting commands, it is up to the |
| 486 | implementation of render() to enable stencil testing with operations |
| 487 | \c KEEP, comparison function \c EQUAL, and a read and write mask of \c 0xFF. |
| 488 | */ |
| 489 | |
| 490 | /*! |
| 491 | \fn const QRegion *QSGRenderNode::RenderState::clipRegion() const |
| 492 | |
| 493 | \return the current clip region or null for backends where clipping is |
| 494 | implemented via stencil or scissoring. |
| 495 | |
| 496 | The \c software backend uses no projection, scissor or stencil, meaning most |
| 497 | of the render state is not in use. However, the clip region that can be set |
| 498 | on the QPainter still has to be communicated since reconstructing this |
| 499 | manually in render() is not reasonable. It can therefore be queried via |
| 500 | this function. The region is in world coordinates and can be passed |
| 501 | to QPainter::setClipRegion() with Qt::ReplaceClip. This must be done before |
| 502 | calling QPainter::setTransform() since the clip region is already mapped to |
| 503 | the transform provided in QSGRenderNode::matrix(). |
| 504 | */ |
| 505 | |
| 506 | /*! |
| 507 | \return pointer to a \a state value. |
| 508 | |
| 509 | Reserved for future use. |
| 510 | */ |
| 511 | void *QSGRenderNode::RenderState::get(const char *state) const |
| 512 | { |
| 513 | Q_UNUSED(state); |
| 514 | return nullptr; |
| 515 | } |
| 516 | |
| 517 | QT_END_NAMESPACE |
| 518 | |