mirror of
https://github.com/Tencent/ncnn.git
synced 2026-10-18 05:36:58 +08:00
nihui
50 lines
1.5 KiB
C++
50 lines
1.5 KiB
C++
// Copyright 2026 Tencent
|
|
// SPDX-License-Identifier: BSD-3-Clause
|
|
|
|
#include "perfutil.h"
|
|
|
|
static void perf_deconvolution(int w, int h, int c, int outch, int kernel, int dilation, int stride, int pad)
|
|
{
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, outch);
|
|
pd.set(1, kernel);
|
|
pd.set(2, dilation);
|
|
pd.set(3, stride);
|
|
pd.set(4, pad);
|
|
pd.set(5, 1);
|
|
pd.set(6, outch * c * kernel * kernel);
|
|
|
|
std::vector<ncnn::Mat> weights(2);
|
|
weights[0] = PerfMat(outch * c * kernel * kernel);
|
|
weights[1] = PerfMat(outch);
|
|
|
|
perf_layer("Deconvolution", pd, weights, PerfMat(w, h, c),
|
|
"out=%d k=%d d=%d s=%d p=%d", outch, kernel, dilation, stride, pad);
|
|
}
|
|
|
|
int main()
|
|
{
|
|
// typical decoder / generator upsampling configurations
|
|
perf_deconvolution(7, 7, 512, 256, 4, 1, 2, 1);
|
|
perf_deconvolution(14, 14, 256, 128, 4, 1, 2, 1);
|
|
perf_deconvolution(28, 28, 128, 64, 4, 1, 2, 1);
|
|
perf_deconvolution(56, 56, 64, 32, 4, 1, 2, 1);
|
|
|
|
// kernel=2 stride=2 (sub-pixel style)
|
|
perf_deconvolution(14, 14, 256, 128, 2, 1, 2, 0);
|
|
perf_deconvolution(28, 28, 128, 64, 2, 1, 2, 0);
|
|
|
|
// kernel=3 stride=1 (no upsampling, dilated)
|
|
perf_deconvolution(14, 14, 256, 256, 3, 1, 1, 1);
|
|
perf_deconvolution(14, 14, 256, 256, 3, 2, 1, 2);
|
|
|
|
// small spatial, large channel
|
|
perf_deconvolution(4, 4, 512, 512, 4, 1, 2, 1);
|
|
|
|
// non-pack4-aligned channel counts
|
|
perf_deconvolution(14, 14, 13, 8, 3, 1, 1, 1);
|
|
perf_deconvolution(14, 14, 64, 13, 4, 1, 2, 1);
|
|
|
|
return 0;
|
|
}
|