here is the inference time, AIP is much worse than DSP, why is that?
,, ,AIP_ub_tf8_timing(2 runs), , , ,DSP_ub_tf8_timing(2 runs), , ,,, ,avg (us),max (us),min (us),runtime,avg (us),max (us),min (us),runtime,,Total Inference Time,3121,3367,2876,CPU|AIP,2366,2418,2314,CPU|DSP
test on SnapDragon 855, with snap benchmark script, here is the json configuration:
{****"Runtimes":["AIP", "DSP"],"Measurements": ["timing"],"CpuFallback": false,"PerfProfile": "high_performance","ProfilingLevel": "detailed","BufferTypes": ["ub_tf8"]}
this is model information
DLC info for/home/suntao/caffescript/snpe-dlc-models/conv_imgS_64_imgC_128_kS_3_kO_128_1layer_rand_quantized.dlcModel Version: N/AModel Copyright:N/A----------------------------------------------------------------------------------------------------------------------------------------------| Id | Name | Type | Inputs | Outputs | Out Dims | Parameters |----------------------------------------------------------------------------------------------------------------------------------------------| 0 | input | data | input | input | 1x64x64x3 | output encoding: min -1, max 0.9922, delta 0.007812, offset -128 bitwidth 8 || | | | | | | input_preprocessing: passthrough || | | | | | | input_type: default || 1 | conv0 | convolutional | input | conv0 | 1x64x64x128 | output encoding: min -15.61, max 16.75, delta 0.1269, offset -123 bitwidth 8 || | | | | | | weight encoding: min 0, max 0.9995, delta 0.00392, offset 0 bitwidth 8 || | | | | | | bias encoding: min 0, max 0.01, delta 3.922e-05, offset 0 bitwidth 8 || | | | | | | padding x: 1 || | | | | | | padding y: 1 || | | | | | | padding mode: zero || | | | | | | stride x: 1 || | | | | | | stride y: 1 || | | | | | | num filters: 128 || | | | | | | kernel: 3x3 || | | | | | | param count: 3k (2.37%) || | | | | | | MACs per inference: 14M (2.29%) || 2 | conv1 | convolutional | conv0 | conv1 | 1x64x64x128 | output encoding: min -6982, max 7612, delta 57.23, offset -122 bitwidth 8 || | | | | | | weight encoding: min 0, max 1, delta 0.003922, offset 0 bitwidth 8 || | | | | | | bias encoding: min 0, max 0.01, delta 3.922e-05, offset 0 bitwidth 8 || | | | | | | padding x: 1 || | | | | | | padding y: 1 || | | | | | | padding mode: zero || | | | | | | stride x: 1 || | | | | | | stride y: 1 || | | | | | | num filters: 128 || | | | | | | kernel: 3x3 || | | | | | | param count: 147k (97.6%) || | | | | | | MACs per inference: 603M (97.7%) |Total parameters: 151168 (0 MB assuming single precision float)Total MACs per inference: 618M (100%)Converter command: snpe-caffe-to-dlc verbose=False encoding=bgr enable_strict_validation=False disable_batchnorm_folding=False input_types=None model_version=None validation_target=[] enable_preprocessing=True input_size=None copyright_file=None input_layers=NoneDLC created with converter version: 1.25.0.287----------------------------------------------------------------------------------------------------------------------------------------------AIP Info:----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| AIP Record Name | nnc_version | hta_driver_version | record_version | record_size | Subnets |----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| aip.metadata0 | 1.2.1.0 | 1.0.0.0 | 1.1.1.0 | 6690 | num_of_subnets: 1 || | | | | | subnet_0: || | | | | | start_layer_Id: 0 || | | | | | end_layer_Id: 2 || | | | | | input_buffers: || | | | | | [name:input],[layer_id:0] || | | | | | output_buffers: || | | | | | [name:conv1],[layer_id:2] || | | | | | target: HTA |