[Perf] Windows/x64: Half and FrozenDictionary Regressions on 1/28/2025 11:31:49 PM +00:00

[performanceautofiler]

Run Information

Name	Value
Architecture	x64
OS	Windows 10.0.22621
Queue	TigerWindows
Baseline	d20c35e05c16dd2ec9cddd4259738cc02228d425
Compare	d9b75154ad360154b863ee01b72fbef94de261b0
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Tests.Perf_Half

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
HalfToSingle - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	1.01 ns	2.25 ns	2.24	0.04	False
HalfToSingle - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	1.01 ns	2.25 ns	2.24	0.04	False

Test Report

Repro

General Docs link: https://github.com/dotnet/performance/blob/main/docs/benchmarking-workflow-dotnet-runtime.md

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Tests.Perf_Half*'

System.Tests.Perf_Half.HalfToSingle(value: 12344)

ETL Files

Histogram

JIT Disasms

System.Tests.Perf_Half.HalfToSingle(value: NaN)

ETL Files

Histogram

JIT Disasms

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

---

Run Information

Name	Value
Architecture	x64
OS	Windows 10.0.22621
Queue	TigerWindows
Baseline	d20c35e05c16dd2ec9cddd4259738cc02228d425
Compare	d9b75154ad360154b863ee01b72fbef94de261b0
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Collections.Perf_Frozen<Int16>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
ToFrozenDictionary - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	50.64 ns	59.55 ns	1.18	0.01	False

Test Report

Repro

General Docs link: https://github.com/dotnet/performance/blob/main/docs/benchmarking-workflow-dotnet-runtime.md

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Collections.Perf_Frozen<Int16>*'

System.Collections.Perf_Frozen<Int16>.ToFrozenDictionary(Count: 4)

ETL Files

Histogram

JIT Disasms

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

[LoopedBard3]

From range: 9b24fb6...21ab780, Half test regressions were caused by #111024 (FYI @EgorBo) and Frozen Dictionary Regression was caused by #111886 (FYI @stephentoub).

Related Regressions:

• [Perf] Windows/x64: 2 Regressions on 1/29/2025 3:52:02 AM +00:00 perf-autofiling-issues#49463 (Only the FrozenDictionary test)
• [Perf] Linux/x64: 3 Regressions on 1/28/2025 11:31:49 PM +00:00 perf-autofiling-issues#49453
• [Perf] Windows/x64: 5 Regressions on 1/28/2025 11:31:49 PM +00:00 perf-autofiling-issues#49445
• [Perf] Linux/x64: 3 Regressions on 1/28/2025 11:31:49 PM +00:00 perf-autofiling-issues#49402
• [Perf] Linux/arm64: 3 Regressions on 1/29/2025 3:52:02 AM +00:00 perf-autofiling-issues#49696
• [Perf] Windows/arm64: 2 Regressions on 1/29/2025 3:52:02 AM +00:00 perf-autofiling-issues#50155

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[xtqqczze]

@tannergooding Is this beneficial? 2 bytes higher in code size, but 0.75 lower PerfScore.

MihuBot/runtime-utils#854

 ; Assembly listing for method System.Half:op_Explicit(float):System.Half (FullOpts)
+       xor      ecx, ecx
+       mov      edx, -1
        vucomiss xmm0, xmm0
-       setp     cl
-       movzx    rcx, cl
-       dec      ecx
+       cmovnp   ecx, edx

Related: #61761 (comment)

Originally posted by @xtqqczze in #111024 (comment)

[stephentoub]

The ToFrozenDictionary one is acceptable. #112298 will help a little, but this is a 9ns regression when constructing the FrozenDictionary, in order to improve the throughput of accessing it. That kind of tradeoff is the main reason this type exists: we're ok spending a bit more at construction in order to optimize throughput of reads.