[null,null,["最后更新时间 (UTC)：2025-07-27。"],[],[],null,["# Analyze vertex formats\n\nYou may diagnose a few possible vertex-related performance problems through the\nuse of frame profiling. Use the **Commands** pane to view all of the draw calls\nyour game performs in a given frame and counts of primitives drawn per draw\ncall. This can give an approximation of the overall number of vertices submitted\nin a single frame.\n**Figure 1.** Frame profiling view for a single `glDrawElements` call, showing 2,718 triangle primitives drawn\n\nVertex attribute compression\n----------------------------\n\nOne common problem your game may face is a large average vertex size. A\nlarge number of vertices submitted with a high average vertex size results in a\nlarge vertex memory read bandwidth when read by the GPU.\n\nTo observe the vertex format for a given draw call, complete the following steps:\n\n1. Select a draw call of interest.\n\n This can be a typical draw call for the scene, a draw call with a large\n number of vertices, a draw call for a complex character model, or some other\n type of draw call.\n2. Navigate to the **Pipeline** pane, and click **IA** for input assembly.\n This defines the vertex format for vertices coming into the GPU.\n\n3. Observe a series of attributes and their formats; for example,\n `R32G32B32_SFLOAT` is a 3-component 32-bit signed float.\n\n**Figure 2.**Input assembly for a draw call, with uncompressed attributes resulting in a vertex size of 56 bytes\n\nFrequently, vertex attributes can be compressed with minimal reduction in the\nquality of the models drawn. In particular, we recommend:\n\n- Compressing vertex position to half-precision 16-bit floats\n- Compressing UV texture coordinates to 16-bit unsigned integer ushorts\n- Compressing the tangent space by encoding normal, tangent, and binormal vectors using quaternions\n\nOther miscellaneous attributes may also be considered for lower-precision types\non a case-by-case basis.\n\nVertex stream splitting\n-----------------------\n\nYou can also investigate whether vertex attribute streams are appropriately\nsplit. On tiled rendering architectures such as mobile GPUs, vertex positions\nare first used in a binning pass to create bins of primitives processed in each\ntile. If vertex attributes are interleaved into a single buffer, all vertex data\nis read into cache for binning, even though only vertex positions are used.\n\nTo reduce vertex read memory bandwidth and improve cache efficiency, and thus\nreduce time spent on the binning pass, vertex data should be split into two\nseparate streams, one for vertex positions, and one for all other vertex\nattributes.\n\nTo investigate whether vertex attributes are appropriately split:\n\n1. Select a draw call of interest, and note the draw call number.\n\n This can be a typical draw call for the scene, a draw call with a large\n number of vertices, a draw call for a complex character model, or some other\n type of draw call.\n2. Navigate to the **Pipeline** pane, and click **IA** for input assembly. This\n defines the vertex format for vertices coming into the GPU.\n\n3. Observe the bindings of your vertex attributes; typically these might\n increase linearly (0, 1, 2, 3, etc.), but this is not always the case.\n Vertex position is typically the first vertex attribute listed.\n\n4. In the **State** pane, find the `LastDrawInfos` and expand the matching draw\n call number. Then, expand the `BoundVertexBuffers` for this draw call.\n\n5. Observe the vertex buffers bound during the given draw call, with indices\n matching the vertex attribute bindings from earlier.\n\n6. Expand the bindings for your draw call's vertex attributes, and expand the\n buffers.\n\n7. Observe the `VulkanHandle` for the buffers, which represent the underlying\n memory that the vertex data sources from. If the `VulkanHandle`s are\n different, this means the attributes originate from different underlying\n buffers. If the `VulkanHandle`s are the same but the offsets are large\n (for example, greater than 100), the attributes may still originate from\n different sub-buffers, but this requires further investigation.\n\n**Figure 3.**Input assembly for a draw call, with the state panel to the right showing that the attributes at binding 0 and 1, vertex position and normal, share a single underlying buffer\n\nFor more detail about vertex stream splitting and how to resolve it on various\ngame engines, see our [blog post](/agi/frame-trace/link-to-Omars-blog-post) on the subject."]]
