MCPcopy Create free account
hub / github.com/Tencent/ncnn / test_convolution_winograd

Function test_convolution_winograd

tests/test_convolution_4.cpp:83–132  ·  view source on GitHub ↗

BF16 winograd path: kernel=3, dilation=1, stride=1, winograd=true, sgemm=false Need num_input>8 || num_output>8 for prefer_winograd winograd43 is the default variant (when neither 63 nor 23 is preferred) winograd63: num_input<64 and specific size ranges (e.g. c=16,outch=16,minwh~25-44) winograd23: large channels with small spatial (e.g. c=512,outch=512,minwh=3-14)

Source from the content-addressed store, hash-verified

source not stored for this graph (policy: none)

Callers 1

mainFunction · 0.85

Calls 1

test_convolutionFunction · 0.70

Tested by

no test coverage detected