mirror of
https://github.com/Tencent/ncnn.git
synced 2026-10-18 21:54:39 +08:00
Lune
70e7bfd608
* [test] reduction: add perf_reduction harness target * [feat] reduction: add x86 optimized Reduction layer Co-authored-by: nihui <nihuini@tencent.com>
119 lines
2.9 KiB
C++
119 lines
2.9 KiB
C++
// Copyright 2026 Tencent
|
|
// SPDX-License-Identifier: BSD-3-Clause
|
|
|
|
#include "perfutil.h"
|
|
|
|
static ncnn::Mat IntArray(int a0)
|
|
{
|
|
ncnn::Mat m(1, (size_t)4u);
|
|
int* p = m;
|
|
p[0] = a0;
|
|
return m;
|
|
}
|
|
|
|
static ncnn::Mat IntArray(int a0, int a1)
|
|
{
|
|
ncnn::Mat m(2, (size_t)4u);
|
|
int* p = m;
|
|
p[0] = a0;
|
|
p[1] = a1;
|
|
return m;
|
|
}
|
|
|
|
static ncnn::Mat IntArray(int a0, int a1, int a2)
|
|
{
|
|
ncnn::Mat m(3, (size_t)4u);
|
|
int* p = m;
|
|
p[0] = a0;
|
|
p[1] = a1;
|
|
p[2] = a2;
|
|
return m;
|
|
}
|
|
|
|
static const char* op_type_name(int op_type)
|
|
{
|
|
static const char* names[] = {"sum", "asum", "sumsq", "mean", "max", "min", "prod", "l1", "l2", "logsum", "logsumexp"};
|
|
if (op_type >= 0 && op_type < 11)
|
|
return names[op_type];
|
|
return "op";
|
|
}
|
|
|
|
static void perf_reduction(const ncnn::Mat& a, int op_type, int keepdims)
|
|
{
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, op_type);
|
|
pd.set(1, 1);
|
|
pd.set(2, 1.f);
|
|
pd.set(4, keepdims);
|
|
|
|
std::vector<ncnn::Mat> weights(0);
|
|
|
|
perf_layer("Reduction", pd, weights, a, "%s axes=all keep=%d", op_type_name(op_type), keepdims);
|
|
}
|
|
|
|
static void perf_reduction(const ncnn::Mat& a, int op_type, int keepdims, int axis)
|
|
{
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, op_type);
|
|
pd.set(1, 0);
|
|
pd.set(2, 1.f);
|
|
pd.set(3, IntArray(axis));
|
|
pd.set(4, keepdims);
|
|
pd.set(5, 1);
|
|
|
|
std::vector<ncnn::Mat> weights(0);
|
|
|
|
perf_layer("Reduction", pd, weights, a, "%s axis=%d keep=%d", op_type_name(op_type), axis, keepdims);
|
|
}
|
|
|
|
static void perf_reduction(const ncnn::Mat& a, int op_type, int keepdims, int axis0, int axis1)
|
|
{
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, op_type);
|
|
pd.set(1, 0);
|
|
pd.set(2, 1.f);
|
|
pd.set(3, IntArray(axis0, axis1));
|
|
pd.set(4, keepdims);
|
|
pd.set(5, 1);
|
|
|
|
std::vector<ncnn::Mat> weights(0);
|
|
|
|
perf_layer("Reduction", pd, weights, a, "%s axes=%d,%d keep=%d", op_type_name(op_type), axis0, axis1, keepdims);
|
|
}
|
|
|
|
static void perf_reduction(const ncnn::Mat& a, int op_type, int keepdims, int axis0, int axis1, int axis2)
|
|
{
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, op_type);
|
|
pd.set(1, 0);
|
|
pd.set(2, 1.f);
|
|
pd.set(3, IntArray(axis0, axis1, axis2));
|
|
pd.set(4, keepdims);
|
|
pd.set(5, 1);
|
|
|
|
std::vector<ncnn::Mat> weights(0);
|
|
|
|
perf_layer("Reduction", pd, weights, a, "%s axes=%d,%d,%d keep=%d", op_type_name(op_type), axis0, axis1, axis2, keepdims);
|
|
}
|
|
|
|
int main()
|
|
{
|
|
perf_reduction(PerfMat(1048576), 0, 0);
|
|
|
|
perf_reduction(PerfMat(1024, 1024), 0, 0, 1);
|
|
perf_reduction(PerfMat(1024, 1024), 0, 0, 0);
|
|
|
|
perf_reduction(PerfMat(1024, 1024), 10, 0, 1);
|
|
perf_reduction(PerfMat(1024, 1024), 10, 0, 0);
|
|
|
|
perf_reduction(PerfMat(56, 56, 64), 3, 1, 1, 2);
|
|
perf_reduction(PerfMat(56, 56, 64), 4, 0, 0);
|
|
|
|
perf_reduction(PerfMat(64, 8, 8, 32), 0, 0, 0, 1, 2);
|
|
perf_reduction(PerfMat(16, 16, 8, 64), 8, 0, 0);
|
|
perf_reduction(PerfMat(16, 16, 8, 64), 1, 0, 2, 3);
|
|
perf_reduction(PerfMat(16, 16, 8, 64), 2, 1, 0, 2, 3);
|
|
|
|
return 0;
|
|
}
|