From 3e31f8820881ca1302f0d2aa664da2a36e1befa2 Mon Sep 17 00:00:00 2001 From: Richard Hsu Date: Tue, 17 May 2022 18:58:17 -0700 Subject: [PATCH] [Tuning] Disable CPU Idle in NNAPI workload with PMQoS CPU DMA Latency To improve the return path latency, we want to keep the CPU at at least WFI state (Idle_1). The PMQos cpu_dma_latency knob prevents the CPU from falling below WFI state. This makes the return path CPU wakeup latency very good. Check with wvw@, the power impact shouldn't be too significant. The average energy cost per inference dropped from 3.85 to 3.47 mJ. The reason why the power number is lower WITH disable-idle is that, due to better latency, we get to run more inferences given the same amount of time. This makes the average power consumption lower. Measurement: MLPerf IC model Latency (ms) Power (mW) Energy/inference (mJ) MLPerf scores Default 1.35 2837 3.85 560 Disable CPU Idle 0.98 3539 3.47 826 https://docs.google.com/presentation/d/1zx7sLkhOClmuRTCrq8-l3N1mZrrv7f-CtcdMuzV0eaI/edit?pli=1#slide=id.g12dd9e50b4b_0_0 Bug: 232183574 Test: MLPerf on Android T. Performance improved. Verified on Perfetto. Change-Id: If067e0851bea0475043ef2127a25ed3a5fdab093 --- powerhint-oriole.json | 15 +++++++++++++++ powerhint-raven.json | 15 +++++++++++++++ 2 files changed, 30 insertions(+) diff --git a/powerhint-oriole.json b/powerhint-oriole.json index 48c0e51..314f634 100644 --- a/powerhint-oriole.json +++ b/powerhint-oriole.json @@ -196,6 +196,15 @@ ], "ResetOnInit": true }, + { + "Name": "PMQoSCpuDmaLatency", + "Path": "/dev/cpu_dma_latency", + "Values": [ + "44", + "1000" + ], + "HoldFd": true + }, { "Name": "CDPreferIdle", "Path": "/proc/vendor_sched/cam_prefer_idle", @@ -1592,6 +1601,12 @@ "Duration": 2000, "Value": "512" }, + { + "PowerHint": "ML_ACC", + "Node": "PMQoSCpuDmaLatency", + "Duration": 2000, + "Value": "44" + }, { "PowerHint": "DEVICE_IDLE", "Node": "RestrictedCpuset", diff --git a/powerhint-raven.json b/powerhint-raven.json index da30e1c..3b9b980 100644 --- a/powerhint-raven.json +++ b/powerhint-raven.json @@ -196,6 +196,15 @@ ], "ResetOnInit": true }, + { + "Name": "PMQoSCpuDmaLatency", + "Path": "/dev/cpu_dma_latency", + "Values": [ + "44", + "1000" + ], + "HoldFd": true + }, { "Name": "CDPreferIdle", "Path": "/proc/vendor_sched/cam_prefer_idle", @@ -1608,6 +1617,12 @@ "Duration": 2000, "Value": "512" }, + { + "PowerHint": "ML_ACC", + "Node": "PMQoSCpuDmaLatency", + "Duration": 2000, + "Value": "44" + }, { "PowerHint": "DEVICE_IDLE", "Node": "RestrictedCpuset",