eden-emu/VulkanMemoryAllocator
14
0
Fork 0
mirror of https://github.com/GPUOpen-LibrariesAndSDKs/VulkanMemoryAllocator.git synced 2025-05-14 16:58:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Refactored virtual allocator. (COMPATIBILITY BREAKING!) Added TLSF algoritym.

Browse source 
Refactored virtual allocator: Added type VmaVirtualAllocation, member VmaVirtualAllocationInfo::offset, changed parameters of vmaVirtualAllocate, vmaVirtualFree, vmaSetVirtualAllocationUserData, vmaGetVirtualAllocationInfo.

Added TLSF algorithm: Added VMA_POOL_CREATE_TLSF_ALGORITHM_BIT, VMA_VIRTUAL_BLOCK_CREATE_TLSF_ALGORITHM_BIT.

Some internal refactoring.

Improved documentation: Grouped API elements into Doxygen modules.

Code mostly by @medranSolus.
This commit is contained in:
Adam Sawicki 2022-01-20 17:00:33 +01:00
parent c5870ad3c2
commit d3a85f0dc3
89 changed files with 7653 additions and 5225 deletions
Download patch file Download diff file Expand all files Collapse all files

24
docs/html/usage_patterns.html
View file

@ -73,7 +73,7 @@ $(function() {
<h1><a class="anchor" id="usage_patterns_common_mistakes"></a>
Common mistakes</h1>
<p ><b>Use of CPU_TO_GPU instead of CPU_ONLY memory</b></p>
<p ><a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a> is recommended only for resources that will be mapped and written by the CPU, as well as read directly by the GPU - like some buffers or textures updated every frame (dynamic). If you create a staging copy of a resource to be written by CPU and then used as a source of transfer to another resource placed in the GPU memory, that staging resource should be created with <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>. Please read the descriptions of these enums carefully for details.</p>
<p ><a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a> is recommended only for resources that will be mapped and written by the CPU, as well as read directly by the GPU - like some buffers or textures updated every frame (dynamic). If you create a staging copy of a resource to be written by CPU and then used as a source of transfer to another resource placed in the GPU memory, that staging resource should be created with <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>. Please read the descriptions of these enums carefully for details.</p>
<p ><b>Unnecessary use of custom pools</b></p>
<p ><a class="el" href="custom_memory_pools.html">Custom memory pools</a> may be useful for special purposes - when you want to keep certain type of resources separate e.g. to reserve minimum amount of memory for them or limit maximum amount of memory they can occupy. For most resources this is not needed and so it is not recommended to create <a class="el" href="struct_vma_pool.html" title="Represents custom memory pool.">VmaPool</a> objects and allocations out of them. Allocating from the default pool is sufficient.</p>
<h1><a class="anchor" id="usage_patterns_simple"></a>
@ -81,23 +81,23 @@ Simple patterns</h1>
<h2><a class="anchor" id="usage_patterns_simple_render_targets"></a>
Render targets</h2>
<p ><b>When:</b> Any resources that you frequently write and read on GPU, e.g. images used as color attachments (aka "render targets"), depth-stencil attachments, images/buffers used as storage image/buffer (aka "Unordered Access View (UAV)").</p>
<p ><b>What to do:</b> Create them in video memory that is fastest to access from GPU using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>.</p>
<p >Consider using <a class="el" href="vk_khr_dedicated_allocation.html">VK_KHR_dedicated_allocation</a> extension and/or manually creating them as dedicated allocations using <a class="el" href="vk__mem__alloc_8h.html#ad9889c10c798b040d59c92f257cae597a3fc311d855c2ff53f1090ef5c722b38f" title="Set this flag if the allocation should have its own memory block.">VMA_ALLOCATION_CREATE_DEDICATED_MEMORY_BIT</a>, especially if they are large or if you plan to destroy and recreate them e.g. when display resolution changes. Prefer to create such resources first and all other GPU resources (like textures and vertex buffers) later.</p>
<p ><b>What to do:</b> Create them in video memory that is fastest to access from GPU using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>.</p>
<p >Consider using <a class="el" href="vk_khr_dedicated_allocation.html">VK_KHR_dedicated_allocation</a> extension and/or manually creating them as dedicated allocations using <a class="el" href="group__group__alloc.html#ggad9889c10c798b040d59c92f257cae597a3fc311d855c2ff53f1090ef5c722b38f" title="Set this flag if the allocation should have its own memory block.">VMA_ALLOCATION_CREATE_DEDICATED_MEMORY_BIT</a>, especially if they are large or if you plan to destroy and recreate them e.g. when display resolution changes. Prefer to create such resources first and all other GPU resources (like textures and vertex buffers) later.</p>
<h2><a class="anchor" id="usage_patterns_simple_immutable_resources"></a>
Immutable resources</h2>
<p ><b>When:</b> Any resources that you fill on CPU only once (aka "immutable") or infrequently and then read frequently on GPU, e.g. textures, vertex and index buffers, constant buffers that don't change often.</p>
<p ><b>What to do:</b> Create them in video memory that is fastest to access from GPU using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>.</p>
<p >To initialize content of such resource, create a CPU-side (aka "staging") copy of it in system memory - <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>, map it, fill it, and submit a transfer from it to the GPU resource. You can keep the staging copy if you need it for another upload transfer in the future. If you don't, you can destroy it or reuse this buffer for uploading different resource after the transfer finishes.</p>
<p ><b>What to do:</b> Create them in video memory that is fastest to access from GPU using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>.</p>
<p >To initialize content of such resource, create a CPU-side (aka "staging") copy of it in system memory - <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>, map it, fill it, and submit a transfer from it to the GPU resource. You can keep the staging copy if you need it for another upload transfer in the future. If you don't, you can destroy it or reuse this buffer for uploading different resource after the transfer finishes.</p>
<p >Prefer to create just buffers in system memory rather than images, even for uploading textures. Use <code>vkCmdCopyBufferToImage()</code>. Dont use images with <code>VK_IMAGE_TILING_LINEAR</code>.</p>
<h2><a class="anchor" id="usage_patterns_dynamic_resources"></a>
Dynamic resources</h2>
<p ><b>When:</b> Any resources that change frequently (aka "dynamic"), e.g. every frame or every draw call, written on CPU, read on GPU.</p>
<p ><b>What to do:</b> Create them using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a>. You can map it and write to it directly on CPU, as well as read from it on GPU.</p>
<p ><b>What to do:</b> Create them using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a>. You can map it and write to it directly on CPU, as well as read from it on GPU.</p>
<p >This is a more complex situation. Different solutions are possible, and the best one depends on specific GPU type, but you can use this simple approach for the start. Prefer to write to such resource sequentially (e.g. using <code>memcpy</code>). Don't perform random access or any reads from it on CPU, as it may be very slow. Also note that textures written directly from the host through a mapped pointer need to be in LINEAR not OPTIMAL layout.</p>
<h2><a class="anchor" id="usage_patterns_readback"></a>
Readback</h2>
<p ><b>When:</b> Resources that contain data written by GPU that you want to read back on CPU, e.g. results of some computations.</p>
<p ><b>What to do:</b> Create them using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a>. You can write to them directly on GPU, as well as map and read them on CPU.</p>
<p ><b>What to do:</b> Create them using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a>. You can write to them directly on GPU, as well as map and read them on CPU.</p>
<h1><a class="anchor" id="usage_patterns_advanced"></a>
Advanced patterns</h1>
<h2><a class="anchor" id="usage_patterns_integrated_graphics"></a>
@ -108,9 +108,9 @@ Detecting integrated graphics</h2>
Direct access versus transfer</h2>
<p >For resources that you frequently write on CPU and read on GPU, many solutions are possible:</p>
<ol type="1">
<li>Create one copy in video memory using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>, second copy in system memory using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a> and submit explicit transfer each time.</li>
<li>Create just a single copy using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a>, map it and fill it on CPU, read it directly on GPU.</li>
<li>Create just a single copy using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>, map it and fill it on CPU, read it directly on GPU.</li>
<li>Create one copy in video memory using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>, second copy in system memory using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a> and submit explicit transfer each time.</li>
<li>Create just a single copy using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a>, map it and fill it on CPU, read it directly on GPU.</li>
<li>Create just a single copy using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>, map it and fill it on CPU, read it directly on GPU.</li>
</ol>
<p >Which solution is the most efficient depends on your resource and especially on the GPU. It is best to measure it and then make the decision. Some general recommendations:</p>
<ul>
@ -120,8 +120,8 @@ Direct access versus transfer</h2>
</ul>
<p >Similarly, for resources that you frequently write on GPU and read on CPU, multiple solutions are possible:</p>
<ol type="1">
<li>Create one copy in video memory using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>, second copy in system memory using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a> and submit explicit tranfer each time.</li>
<li>Create just single copy using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a>, write to it directly on GPU, map it and read it on CPU.</li>
<li>Create one copy in video memory using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>, second copy in system memory using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a> and submit explicit tranfer each time.</li>
<li>Create just single copy using <a class="el" href="group__group__alloc.html#ggaa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a>, write to it directly on GPU, map it and read it on CPU.</li>
</ol>
<p >You should take some measurements to decide which option is faster in case of your specific resource.</p>
<p >Note that textures accessed directly from the host through a mapped pointer need to be in LINEAR layout, which may slow down their usage on the device. Textures accessed only by the device and transfer operations can use OPTIMAL layout.</p>