summaryrefslogtreecommitdiffstats
path: root/src/gui/rhi
Commit message (Collapse)AuthorAgeFilesLines
* rhi: vk: Prevent WAW hazard reported from sync validation with offscreen rtLaszlo Agocs2021-01-062-4/+45
| | | | | | | | | | | | ...when a QVkRenderBuffer is used as depth-stencil (i.e. a "throwaway" depth-stencil buffer that is a lazily allocated / transient VkImage under the hood). Such a resource is not tracked by the QRhiPassResourceTracker, so do an explicit barrier in beginPass instead of having it generated by the tracker. Fixes: QTBUG-89764 Change-Id: Ice794f44342175f712ea56f450270cbb8929f516 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Reset tracked state upon a buffer update or readbackLaszlo Agocs2021-01-061-48/+70
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | ...encountered in the command list. Move all, previously local, tracking variables into a struct. This allows creating helper functions to reduce error-prone repetition in the executeCommandBuffer() function body. The only real change in the patch is in the handling of Command::BufferSubData and Command::GetBufferSubData: here, instead of calling glBindBuffer directly, use a helper function that also resets the relevant state tracking variables. A subsequent Command::BindVertexBuffer or BindIndexBuffer will therefore correctly rebind the appropriate buffers. This is particularly relevant with certain command stream patterns exercised by some Qt Quick 3D scenes: - A View3D renders a mesh, - another View3D has some 2D Qt Quick content, as well as a model with the same mesh. When both View3Ds use the default Offscreen render mode, the resulting command list consists of segments along the lines of: 1. prepare resources for first View3D 2. render content for first View3D - this binds the vertex and index buffers for the mesh (state is tracked; all 1-4 steps are within the same command list, processed by a single call to executeCommandBuffer()) 3. prepare the content for the "inline" 2D Qt Quick scene - this may update vertex and index buffers, that may lead to adding BufferSubData commands to the list (tracked state (last vertex/index buffer) may need invalidation/updating - and that's where our problem lies) 4. the second View3Ds 3D content is rendered: a model with the same mesh as the last (Quick)3D draw call, so same vertex and index buffers. If #3 did not invalidate and/or update the tracked state, the glBindBuffer calls are (incorrectly) skipped. Fixes: QTBUG-89780 Pick-to: 6.0 Change-Id: Icc933252f3993b8727d192e7ba4aa0f842bab51e Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: vk: Fix DS WAW hazard reported from sync validationLaszlo Agocs2021-01-041-10/+23
| | | | | | | | | | | | | | Running with synchronization validation (SDK 1.2.154) we get: Hazard WRITE_AFTER_WRITE vs. layout transition in subpass 0 for attachment 1 aspect depth during load with loadOp VK_ATTACHMENT_LOAD_OP_CLEAR Attempt to fix this with an appropriate subpass dependency. Fixes: QTBUG-89760 Pick-to: 6.0 Change-Id: I67e15a09aab564f63aab646761ccdb1f8b82111d Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* Add QRhiTextureSubresourceUploadDescription QByteArray constructorJonas Karlsson2020-12-292-2/+14
| | | | | Change-Id: I1be4ef1a15c4049f19d0401d6a1a62b3117180a0 Reviewed-by: Eirik Aavitsland <eirik.aavitsland@qt.io>
* Prevent copy in texture upload with invalid inputLaszlo Agocs2020-12-161-2/+4
| | | | | | | | | | Skip the memcpy when src is null. Also silences a Codechecker warning. Pick-to: 6.0 Change-Id: I5042d725000cb6dff6864408fa9ed9e0ca35145a Reviewed-by: Christian Strømme <christian.stromme@qt.io>
* Replace discouraged Q_MOVABLE_TYPE by Q_RELOCATABLE_TYPEAndreas Buhr2020-11-309-40/+40
| | | | | | | | | | | | | | Q_MOVABLE_TYPE was conceived before C++ had move semantics. Now, with move semantics, its name is misleading. Q_RELOCATABLE_TYPE was introduced as a synonym to Q_MOVABLE_TYPE. Usage of Q_MOVABLE_TYPE is discouraged now. This patch replaces all usages of Q_MOVABLE_TYPE by Q_RELOCATABLE_TYPE in QtBase. As the two are synonymous, this patch should have no impact on users. Pick-to: 6.0 Change-Id: Ie653984363198c1aeb1f70f8e0fa189aae38eb5c Reviewed-by: Lars Knoll <lars.knoll@qt.io> Reviewed-by: Andrei Golubev <andrei.golubev@qt.io>
* Fix typo in rhiAntti Määttä2020-11-271-1/+1
| | | | | | Pick-to: 6.0 Change-Id: I2e920e8e68b7010b565c422bb26708d172d345e8 Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
* RHI: Fix attribute instancing cleanup for OpenGL ESPaul Olav Tvete2020-11-271-0/+9
| | | | | | | | | | | | | The previous attribute instancing patch (6493b93) performed the cleanup too late for the case where the command buffer ends with an EndFrame command and core profile is used. Resetting the attribute divisors needs to be done before the vertex array is unbound. Otherwise the state will be wrong at the start of the next call to executeCommandBuffer(), which is normally the start of the next frame. Change-Id: Ic76695b4d334ed1c1e816e747417d957c387a88b Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
* RHI: Reset attribute instancing properly for OpenGL ESPaul Olav Tvete2020-11-261-3/+24
| | | | | | | | | | | | Enabling instanced drawing will set glVertexAttribDivisor to 1 for the PerInstance attributes. This is of course a persistent state, because GL, so it will apply to all subsequent draw calls that uses attributes with the same location number, introducing weird and wonderful randomness. Therefore, make sure we set the divisor back to 0 for non-instanced attributes. Change-Id: I2d9115369fa24e8d57396d2a5f88d1435fe98971 Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
* rhi: gl: Do not bother with glDrawBuffers(GL_BACK)Laszlo Agocs2020-11-021-4/+0
| | | | | | | | | | | | | | | | | It's per framebuffer and the color target for the default fbo is never altered. So no need for this call. This is useful because it avoids having to deal with the controversy around GL_BACK (where the GL, but not the ES, spec claims some quite nonsensical things about GL_BACK not being accepted by glDrawBuffers (in the name of stereo support?), nevermind GL_BACK being the default value...), and while what we have in place now seems to be ok with many implementations, it may generate a GL error with others - so just remove the call in order not have to think about this legacy nonsense) Task-number: QTBUG-88070 Change-Id: I051c60a3041865a3434801b57da2b7acb518b8fe Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* QRhiBackendCommandList: avoid new[]/delete mismatchFabian Kosmale2020-10-191-1/+1
| | | | | | Change-Id: Iac2645ebd1d42753817078f194ba61520f5f70c9 Reviewed-by: Volker Hilsheimer <volker.hilsheimer@qt.io> Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
* rhi: Further reduce copying in the command bufferLaszlo Agocs2020-10-167-228/+170
| | | | | Change-Id: I2e2ff5f4b8aa91d0accb01108a5199b98c371455 Reviewed-by: Christian Strømme <christian.stromme@qt.io>
* rhi: Unroll the res.binding factory functionsLaszlo Agocs2020-10-161-11/+59
| | | | | | | | | Trading typing in a few lines of straightforward code for a full sizeof(QRhiShaderResourceBinding::Data) copy (with less fortunate compilers that do not manage to optimize it away) is a bad deal. Change-Id: If227320f8a9598508c1fe02c4f56e6d75f0a787b Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Fix up vertex inputs with matricesLaszlo Agocs2020-10-143-6/+31
| | | | | | | | | | | | In order to prevent too much voodoo in backends like D3D11, the input layout is expected to specify the slice index for vecX that are part of an unrolled matrix. Also deoptimize the instancing manual test to exercise a matrix too instead of just vectors. Change-Id: If2dcbcbc483645ce2420b2f87dda765b95da6e80 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Add missing \internal doc commandsTor Arne Vestbø2020-10-121-0/+3
| | | | | | Pick-to: 5.15 Change-Id: I09a69a04c6fcd406602e797cbeec59216fecd5d8 Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
* rhi: Make the new direct buffer update available for non-uniform buffersLaszlo Agocs2020-10-1212-41/+69
| | | | | | | | | | | | | | | | | The original restriction to UniformBuffer was due to the GL backend where there is no GL buffer object for QRhiBuffers with usage UniformBuffer. However, we can still implement this for cases when there is a true GL buffer object underneath. With other backends it should all work as-is already. This becomes useful when one has buffers with usage Vertex that need full updates every frame. (f.ex. instance data) Unfortunately this involves renaming the function. But while at it, add an autotest case as well. Change-Id: Iff59e4509a8bae06654cc92fe8428bd79eb012fb Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: d3d: Save a few cycles in bindShaderResourcesLaszlo Agocs2020-10-113-121/+144
| | | | | Change-Id: Ib11ea33eb695d4599f4f040415d497aaf19cb15b Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: d3d: Calculate dynamic offset usage upfrontLaszlo Agocs2020-10-112-9/+17
| | | | | | | Also remove an unnecessary buffer type test. Change-Id: I6f8312f03d3fe1729fa9b2eb95f6bfbeeccd0d4f Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: vk: Avoid shadowing local variableLaszlo Agocs2020-10-111-5/+6
| | | | | Change-Id: I7ad14ff05a1ff10113f5d3985662970d74e68ead Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: vk: Skip buffer/texture registration on read-after-readLaszlo Agocs2020-10-111-6/+19
| | | | | Change-Id: I22027cfc227d3c09e446d193e6b2903b8df34eb2 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: vk: Pre-calculate the flags for dyn.offsetLaszlo Agocs2020-10-114-48/+54
| | | | | | | ...instead of doing a loop in setShaderResources() just for this. Change-Id: Iac8d4517783967c6b8bca4926cceca918f7dcdec Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: d3d: Reduce the size of the Command structLaszlo Agocs2020-10-115-23/+36
| | | | | | | Copied by value so the size matters. Change-Id: I17eae99212801a4fb390a0e298b361123644d17d Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Move some code out of the inner loopsLaszlo Agocs2020-10-111-4/+4
| | | | | Change-Id: I96c71ef45b93ab95250a2a3f0e9c10e8bc86d650 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Add support for full, direct buffer updatesLaszlo Agocs2020-10-1113-6/+160
| | | | | Change-Id: I02c1f8c32c08d39cde9845d20ba8b02541d9d325 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Silence incorrect compiler warnings about uninitialized useLaszlo Agocs2020-10-081-3/+3
| | | | | Change-Id: I09bd22f3ccb1231ad4ef123f581459ef5c1ac37c Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
* rhi: gl: Drop the unused WRAP_R memberLaszlo Agocs2020-10-082-5/+0
| | | | | | | | If we do not use it for anything, do not spend time on comparing and copying it either. Change-Id: I3632792847010e5b73c48eb1ca88f5806f3ccd45 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Fix missing texture bindingsLaszlo Agocs2020-10-081-5/+7
| | | | | | | | The block to update the cached state needs to be run also when the texture parameters have changed. Change-Id: Ifa80c142c6ff31c95a718a1a900ff3db0bbfe4bc Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: vk: Reduce memcopies with full dynamic buffer updatesLaszlo Agocs2020-10-051-1/+4
| | | | | Change-Id: Ifc754a639fcdb45a1b1b8721393de9356504342f Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Add a note to generateMips docsLaszlo Agocs2020-10-021-0/+7
| | | | | Change-Id: Ib07eef8944367948338075832452fdae6df6078a Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: vk: Alter command buffer strategyLaszlo Agocs2020-10-023-131/+158
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Reports on the Web have it that there's nothing guaranteeing a driver does any actual freeing of resources then doing vkFreeCommandBuffer for a command buffer from a command pool that does not have VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT, thus leading to continuously growing resource usage with our current allocate/free pattern. It could be that this is the source of out of memory problems we are seeing on some Android devices. Instead of just going straight for said command pool flag and doing ResetCommandBuffer for the command buffers individually, take one step further and use per-slot (slot being 0 or 1 if QVK_FRAMES_IN_FLIGHT is 2) command pools. The current pool is reset in each beginFrame/beginOffscreenFrame, moving all allocated command buffers to the initial state (while other command buffers with the other pool are not affected). This may be (while impossible to tell from just guessing based on the spec) our best approach to command buffer allocation since a Vulkan implementation can likely just use some simple per pool allocator, knowing that we never want to free or reset individual command buffers, but we rather only reset the whole pool at once. The option of importing an existing VkCommandPool when creating the QRhi instance is now gone, but there was probably no point in offering that in the first place. When it comes to VK_COMMAND_POOL_RESET_RELEASE_RESOURCES_BIT it will not be set unless releaseCachedResources() (in Qt Quick this is hooked into QQuickWindow::releaseResources()) was called. What this does in practice is unknown, but have an option to set it now and then if the application really wants. While we are at it, rename secondaryCbs to activeSecondaryCbStack to indicate what it really is. (it's a stack as each call to beginExternal() pushes a new one, while each endExternal() pops) Change-Id: I2e5c1fad26d794e1f56c778e38f750998d706d84 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Best uniform is no uniformLaszlo Agocs2020-10-022-4/+84
| | | | | | | | | | | | | | | | | Artificial tests with tens of thousands of meshes drawn with the same program (graphics pipeline) and almost the same uniform values (the difference being in the normal and model and view matrices) do benefit - strongly depending on the GL implementation probably - from doing fewer glUniformNf[v] calls. Building on the fact that uniform location values will typically be an int value starting at 0 (not guaranteed of course, we just skip the smartness in that case), we can dedicate a small 16K block to keep track of float/vec3/vec3/vec4 values in the graphics and compute pipeline objects. Change-Id: I217c31ccdeb511b3e8b8286078d7fbde399c8e3b Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Stop flooding with glVertexAttribPointersLaszlo Agocs2020-10-021-6/+26
| | | | | | | | | | | Artificial tests like the BechmarkDemo are full of drawing the exact same mesh multiple times in a row. These do not need respecifying anything related to vertex or index data. Also move the buffer binding out from the loop. Change-Id: I0f27a39fecebc7ca8e1fa635c63819f116867e19 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Skip rebinding the same index buffer in the same passLaszlo Agocs2020-10-021-1/+5
| | | | | | | | Already doing the same for GL_ARRAY_BUFFER, follow suit with GL_ELEMENT_ARRAY_BUFFER. Change-Id: I0cf44480abb3a60a44b316af2c09b2235f9f65e7 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Make per-frame flags per-pass where appropriateLaszlo Agocs2020-10-0213-124/+196
| | | | | | | | | | | | | | | | | | | | ExternalContentsInPass becomes a per-pass flag now. Why is this beneficial? Because while Qt Quick has no choice for its render pass, not being able to guess if the application wants to do some native rendering in there, Quick 3D's render passes, all the ones that are under Quick3D's control, do not have native rendering from the application in them, and so using secondary command buffers with Vulkan is not necessary. Introduce something similar for compute and OpenGL. By knowing that none of the resources used in a pass are used with a compute pass (e.g. because we know that there are no compute passes at all) a small amount of time can be saved by skipping tracking buffers and textures because the only purpose of said tracking is to generate barriers that are relevant only to compute. Change-Id: I0eceb4774d87803c73a39db527f5707a9f4d75c1 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Attempt to reduce superfluous texture unit bindingsLaszlo Agocs2020-10-022-16/+58
| | | | | Change-Id: Iccfbe5fe8688245a5ff3f7d46a173638b362a3c8 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Manage buffer data without QBALaszlo Agocs2020-09-309-63/+119
| | | | | | | | Use a simple and straightforward container that only does what we need here. Change-Id: I1a81b53a58bc91d533e3d7df5471a1362046825d Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: vk: Do not copy the entire BufferOp struct for host writesLaszlo Agocs2020-09-292-7/+11
| | | | | | | | Take only the three things we need. Otherwise we waste time on copying data that is not even relevant to buffer updates at all. Change-Id: I5ed6ae647e23c6f1d0f5f1d973bead2e008f06cc Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Skip nice but not essential uniform checks in releaseLaszlo Agocs2020-09-291-0/+2
| | | | | Change-Id: Ic258f128d691a221670d930d2bfe90cda74a4c75 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Drop QBitArray usageLaszlo Agocs2020-09-292-8/+13
| | | | | Change-Id: I4ae92e6c8c91111a4593c51ee05443b3bc806c35 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Improve layout compatibility test performanceLaszlo Agocs2020-09-298-9/+36
| | | | | | | | | Also bump the non-heap buffer size in the binding list to 16, in order to accommodate complex Quick3D materials with many associated texture maps. Change-Id: Id190e5f8304f5941cffc41a2605fce45dfeb72f0 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Pre-sort uniform metadata based on offsetLaszlo Agocs2020-09-291-7/+11
| | | | | | | | | Because we will iterate through this list and issue a memcpy for each entry. Better to keep it sorted based on offset to be more cache friendly. Change-Id: Ie9dcb259e9a543937cbdcdea85aec9eb92dba1b1 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Switch command lists to QVLALaszlo Agocs2020-09-253-3/+3
| | | | | Change-Id: Ic163533eee973f0ee2f3e2efe25390caa57dd659 Reviewed-by: Paul Olav Tvete <paul.tvete@qt.io>
* rhi: gl: get rid of a wrapper QByteArray where a raw ptr sufficesLaszlo Agocs2020-09-251-3/+2
| | | | | | | | This is a visible performance increase actually, given this is on a quite hot chode path. Change-Id: I165bfa0d4e490f80078551efb4bf2fe8e72d3596 Reviewed-by: Paul Olav Tvete <paul.tvete@qt.io>
* rhi: Skip comparing to this in isCompatible testsLaszlo Agocs2020-09-252-0/+9
| | | | | Change-Id: Ie1855f992315f6e02986016a254127cec0ac94c8 Reviewed-by: Paul Olav Tvete <paul.tvete@qt.io>
* rhi: gl: d3d: Do not use QByteArray where there is no pointLaszlo Agocs2020-09-254-14/+17
| | | | | | | | | The real benefit in practice is getting rid of profilers going wild about QByteArray::begin() and similar. The actual perf. gain is fairly small. Change-Id: I25e4c762f303b41f0aaf914bc62751f3265d3cf8 Reviewed-by: Paul Olav Tvete <paul.tvete@qt.io>
* rhi: Drop the zeroing out of resource binding structsv6.0.0-alpha1Laszlo Agocs2020-09-252-10/+2
| | | | | | | | | | This has a non-insignificant cost (since the size is at least 260 bytes) in renderers that construct a lot of these objects per frame. After changing the qHash implementation to be lighter we no longer rely on qHashBits, so zeroing it is not essential anymore. Remove it. Change-Id: If1fbd8dfd46fb3f0e9ea5a3c4794d24b8aeac1ba Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: vulkan: Fix mipmap generation for cubemapsLaszlo Agocs2020-09-254-80/+81
| | | | | Change-Id: Ia1aab06214be802aaabc97ffefa28947e11148e3 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: gl: Enable seamless cubemap filteringLaszlo Agocs2020-09-231-0/+10
| | | | | | | | ...when available. This is mandatory in order to match results with other APIs where this is always on. Change-Id: I3b8ef2cb806a7297c83e60a3eb1fa71b50780dbe Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Reuse the data in buffer ops in res.update batchesLaszlo Agocs2020-09-236-25/+60
| | | | | | | | Because having profilers bombarded with mallocs (due to creating deep copy QByteArrays) is not nice. Change-Id: I848f41f3465d6dc2a58a193cc863495aacf13d79 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
* rhi: Reuse buffer/texture op entries in res.update batchesLaszlo Agocs2020-09-237-36/+125
| | | | | | | | More of an enabler for reusing the data in the individual entries since not clearing the QVLA does not give us much on its own. Change-Id: Ief9761f75382c3373cc2bc7b866eb59fdd8b3277 Reviewed-by: Andy Nichols <andy.nichols@qt.io>
generated by cgit v1.2.3 (git 2.25.1) at 2024-05-08 20:18:55 +0000