MCPcopy Create free account
hub / github.com/DeepRec-AI/DeepRec / ThresholdRatio

Method ThresholdRatio

tensorflow/lite/experimental/ruy/tune.cc:102–124  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

100}
101
102float TuningResolver::ThresholdRatio() {
103 // Empirically (see :tune_tool) determined threshold to distinguish in-order
104 // Cortex-A53/A55 cores from out-of-order Cortex-A57/A73/A75/A76 cores. Based
105 // on these experimental results, which were obtained with much lower
106 // (kLoopIters=1000, kRepeats=1) so as to make them resilient to noise, we
107 // have:
108 //
109 // CPU core type | in/out of order | observed ratio
110 // --------------+-----------------+-----------------------------------------
111 // Cortex-A53 | in-order | 0.32 -- 0.329
112 // Cortex-A55 | in-order | 0.319 -- 0.325
113 // Cortex-A55r1 | in-order | 0.319 -- 0.325
114 // Cortex-A57 | out-of-order | 0.99 -- 1.01
115 // Cortex-A73 | out-of-order | 0.922 -- 0.927
116 // Cortex-A75 | out-of-order | 0.921 -- 0.93
117 // Cortex-A76 | out-of-order | 1
118 // Kryo (pixel1) | out-of-order | 0.73 -- 0.76
119 //
120 // Thus the allowable range for the threshold is [0.35 .. 0.70].
121 // We pick a value closer to the upper bound because really any out-of-order
122 // CPU should by definition produce a ratio close to 1.
123 return 0.65f;
124}
125
126Tuning TuningResolver::ResolveNow() {
127 const bool is_probably_inorder = EvalRatio() < ThresholdRatio();

Callers 1

QueryMethod · 0.80

Calls

no outgoing calls

Tested by

no test coverage detected