2 Commits

Author	SHA1	Message	Date
Ben Clayton	ba1a8f8d05	optimization: BlockAllocator: Actually allocate in blocks ... Instead of hitting the heap for each and every call to Create() Significantly improves performance for heavy loads. Slight performance loss for lighter loads. A: base.bench B: new.bench Test name | Δ (A → B) | % (A → B) --------------------------------------+--------------+----------- GenerateSPIRV/"simple_fragment.wgsl" | 27.021µs | +6.4% GenerateMSL/"simple_compute.wgsl" | 35.592µs | +6.1% GenerateMSL/"simple_vertex.wgsl" | 37.64µs | +5.5% GenerateHLSL/"simple_fragment.wgsl" | 42.145µs | +5.2% GenerateGLSL/"simple_fragment.wgsl" | 31.506µs | +4.9% GenerateHLSL/"simple_vertex.wgsl" | 38.843µs | +4.7% GenerateMSL/"simple_fragment.wgsl" | 29.977µs | +4.5% GenerateSPIRV/"simple_vertex.wgsl" | 19.882µs | +4.2% GenerateGLSL/"simple_vertex.wgsl" | 24.702µs | +3.7% GenerateSPIRV/"simple_compute.wgsl" | 17.652µs | +3.2% GenerateHLSL/"simple_compute.wgsl" | 26.826µs | +2.7% GenerateGLSL/"simple_compute.wgsl" | 11.952µs | +1.8% ParseWGSL/"particles.wgsl" | -104.83µs | -4.2% GenerateMSL/"particles.wgsl" | -1.079243ms | -9.4% GenerateSPIRV/"particles.wgsl" | -1.012483ms | -9.4% GenerateGLSL/"particles.wgsl" | -3.522106ms | -9.5% GenerateHLSL/"particles.wgsl" | -1.849666ms | -10.6% Issue: tint:1383 Change-Id: Ib691328538c597c06a75dfba392c99d2afbd5442 Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/76961 Reviewed-by: Antonio Maiorano <amaiorano@google.com> Kokoro: Kokoro <noreply+kokoro@google.com> Commit-Queue: Ben Clayton <bclayton@google.com>	2022-01-22 20:32:38 +00:00
Ben Clayton	587f387fd9	Add tint::BlockAllocator<T> ... A container and allocator of objects of (or deriving from) the template type `T`. Objects are allocated by calling Create(), and are owned by the BlockAllocator. When the BlockAllocator is destructed, all constructed objects are automatically destructed and freed. Objects held by the BlockAllocator can be iterated over using a View or ConstView. Use this to hold the ast::Nodes in the ast::Module This is called BlockAllocator as it can be optimized to hold objects in contiguous memory blocks, which will improve cache coherencey. Currently BlockAllocator is a straight port of the vector-of-unique-ptr, taken from ast::Module. Change-Id: I4bf4d298aec3c70d2ddf833e2f168416cbb024c0 Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/38001 Reviewed-by: dan sinclair <dsinclair@chromium.org> Commit-Queue: Ben Clayton <bclayton@google.com>	2021-01-21 15:30:10 +00:00