mirror of
https://github.com/Tencent/ncnn.git
synced 2026-10-18 21:54:39 +08:00
nihui
41f031e045
* adjust tests * opt shader coverage op * update gpu_coverage.cpp atomicor coalesce
319 lines
9.3 KiB
C++
319 lines
9.3 KiB
C++
// Copyright 2021 Tencent
|
|
// SPDX-License-Identifier: BSD-3-Clause
|
|
|
|
#include "testutil.h"
|
|
|
|
#include "layer_type.h"
|
|
|
|
#include <limits.h>
|
|
|
|
#define OP_TYPE_MAX 11
|
|
|
|
static int op_type = 0;
|
|
|
|
static std::vector<int> IntArray(int a0)
|
|
{
|
|
std::vector<int> m(1);
|
|
m[0] = a0;
|
|
return m;
|
|
}
|
|
|
|
static std::vector<int> IntArray(int a0, int a1)
|
|
{
|
|
std::vector<int> m(2);
|
|
m[0] = a0;
|
|
m[1] = a1;
|
|
return m;
|
|
}
|
|
|
|
static void print_int_array(const std::vector<int>& a)
|
|
{
|
|
fprintf(stderr, "[");
|
|
for (size_t i = 0; i < a.size(); i++)
|
|
{
|
|
fprintf(stderr, " %d", a[i]);
|
|
}
|
|
fprintf(stderr, " ]");
|
|
}
|
|
|
|
static int test_reduction(const ncnn::Mat& _a, float coeff, int keepdims)
|
|
{
|
|
ncnn::Mat a = _a;
|
|
if (op_type == 9 || op_type == 10)
|
|
{
|
|
// value must be positive for logsum and logsumexp
|
|
Randomize(a, 0.001f, 2.f);
|
|
}
|
|
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, op_type);
|
|
pd.set(1, 1); // reduce_all
|
|
pd.set(2, coeff);
|
|
pd.set(4, keepdims);
|
|
|
|
std::vector<ncnn::Mat> weights(0);
|
|
|
|
int ret = test_layer("Reduction", pd, weights, a, 0.001);
|
|
if (ret != 0)
|
|
{
|
|
fprintf(stderr, "test_reduction failed a.dims=%d a=(%d %d %d %d) op_type=%d coeff=%f keepdims=%d reduce_all=1\n", a.dims, a.w, a.h, a.d, a.c, op_type, coeff, keepdims);
|
|
}
|
|
|
|
return ret;
|
|
}
|
|
|
|
static int test_reduction(const ncnn::Mat& _a, float coeff, int keepdims, const std::vector<int>& axes_array)
|
|
{
|
|
ncnn::Mat a = _a;
|
|
if (op_type == 9 || op_type == 10)
|
|
{
|
|
// value must be positive for logsum and logsumexp
|
|
Randomize(a, 0.001f, 2.f);
|
|
}
|
|
|
|
ncnn::Mat axes(axes_array.size());
|
|
{
|
|
int* p = axes;
|
|
for (size_t i = 0; i < axes_array.size(); i++)
|
|
{
|
|
p[i] = axes_array[i];
|
|
}
|
|
}
|
|
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, op_type);
|
|
pd.set(1, 0); // reduce_all
|
|
pd.set(2, coeff);
|
|
pd.set(3, axes);
|
|
pd.set(4, keepdims);
|
|
pd.set(5, 1); // fixbug0
|
|
|
|
std::vector<ncnn::Mat> weights(0);
|
|
|
|
int ret = test_layer("Reduction", pd, weights, a, 0.001);
|
|
if (ret != 0)
|
|
{
|
|
fprintf(stderr, "test_reduction failed a.dims=%d a=(%d %d %d %d) op_type=%d coeff=%f keepdims=%d", a.dims, a.w, a.h, a.d, a.c, op_type, coeff, keepdims);
|
|
fprintf(stderr, " axes=");
|
|
print_int_array(axes_array);
|
|
fprintf(stderr, "\n");
|
|
}
|
|
|
|
return ret;
|
|
}
|
|
|
|
static int test_reduction_axes(const ncnn::Mat& a)
|
|
{
|
|
// axis geometry and output rank are independent of coefficient scaling
|
|
int ret = test_reduction(a, 1.f, 0) || test_reduction(a, 2.f, 1);
|
|
if (ret != 0)
|
|
return ret;
|
|
const int axes1[][4] = {{0, -1, -1, -1}};
|
|
const int axes2[][4] = {{0, -1, -1, -1}, {1, -1, -1, -1}, {0, 1, -1, -1}};
|
|
const int axes3[][4] = {
|
|
{0, -1, -1, -1}, {1, -1, -1, -1}, {2, -1, -1, -1}, {0, 1, -1, -1}, {0, 2, -1, -1}, {1, 2, -1, -1}, {0, 1, 2, -1}
|
|
};
|
|
const int axes4[][4] = {
|
|
{0, -1, -1, -1}, {1, -1, -1, -1}, {2, -1, -1, -1}, {3, -1, -1, -1}, {0, 1, -1, -1}, {0, 2, -1, -1}, {0, 3, -1, -1}, {1, 2, -1, -1}, {1, 3, -1, -1}, {2, 3, -1, -1}, {0, 1, 2, -1}, {0, 1, 3, -1}, {0, 2, 3, -1}, {1, 2, 3, -1}, {0, 1, 2, 3}
|
|
};
|
|
const int(*axis_sets)[4] = a.dims == 4 ? axes4 : a.dims == 3 ? axes3 : a.dims == 2 ? axes2 : axes1;
|
|
const int counts[] = {1, 3, 7, 15};
|
|
for (int i = 0; i < counts[a.dims - 1]; i++)
|
|
{
|
|
std::vector<int> axes;
|
|
for (int j = 0; j < 4 && axis_sets[i][j] >= 0; j++)
|
|
axes.push_back(axis_sets[i][j]);
|
|
ret = test_reduction(a, 1.f, 0, axes) || test_reduction(a, 2.f, 1, axes);
|
|
if (ret != 0)
|
|
return ret;
|
|
}
|
|
return 0;
|
|
}
|
|
|
|
static int test_reduction_arithmetic()
|
|
{
|
|
// each operation exercises every axis kernel and both output layouts
|
|
return test_reduction_axes(RandomMat(127))
|
|
|| test_reduction_axes(RandomMat(19, 15))
|
|
|| test_reduction_axes(RandomMat(7, 9, 12))
|
|
|| test_reduction_axes(RandomMat(7, 8, 9, 12))
|
|
|| test_reduction(RandomMat(3, 1, 5, 1), 2.f, 0, IntArray(1, 3));
|
|
}
|
|
|
|
static int test_reduction_vector_boundaries()
|
|
{
|
|
// lengths below, at and above SIMD widths retain vector and scalar loops
|
|
const int sizes[] = {1, 3, 4, 7, 8, 15, 16, 17, 31, 32, 33, 128};
|
|
for (size_t i = 0; i < sizeof(sizes) / sizeof(sizes[0]); i++)
|
|
{
|
|
ncnn::Mat a = RandomMat(sizes[i]);
|
|
ncnn::Mat b = RandomMat(sizes[i], 3);
|
|
int ret = test_reduction(a, 1.f, 0)
|
|
|| test_reduction(a, 2.f, 1)
|
|
|| test_reduction(b, 1.f, 0)
|
|
|| test_reduction(b, 2.f, 1, IntArray(1));
|
|
if (ret != 0)
|
|
return ret;
|
|
}
|
|
return 0;
|
|
}
|
|
|
|
static int test_reduction_geometry()
|
|
{
|
|
// singleton ranks, packed transfers and a large workgroup reduction
|
|
return test_reduction_axes(RandomMat(1, 3, 1, 5))
|
|
|| test_reduction_axes(RandomMat(3, 1, 5, 1))
|
|
|| test_reduction_axes(RandomMat(1, 1, 1, 1))
|
|
|| test_reduction_axes(RandomMat(1, 3, 5))
|
|
|| test_reduction_axes(RandomMat(3, 1, 5))
|
|
|| test_reduction_axes(RandomMat(1, 5))
|
|
|| test_reduction_axes(RandomMat(5, 1))
|
|
|| test_reduction_axes(RandomMat(3, 4, 5, 13))
|
|
|| test_reduction(RandomMat(5, 6, 7, 24), 1.f, 0)
|
|
|| test_reduction(RandomMat(17, 12), 1.f, 0)
|
|
|| test_reduction(RandomMat(5, 7, 24), 2.f, 1, IntArray(2));
|
|
}
|
|
|
|
static int test_reduction_numeric_boundaries()
|
|
{
|
|
// zero L2 results exercise the subnormal square-root guard
|
|
op_type = 8;
|
|
ncnn::Mat a(33);
|
|
ncnn::Mat b(3, 2, 4);
|
|
a.fill(0.f);
|
|
b.fill(0.f);
|
|
int ret = test_reduction(a, 1.f, 0)
|
|
|| test_reduction(b, 2.f, 1, IntArray(2));
|
|
if (ret != 0)
|
|
return ret;
|
|
|
|
// negative axes preserve each operation and both output layouts
|
|
for (op_type = 0; op_type < OP_TYPE_MAX; op_type++)
|
|
{
|
|
ret = test_reduction(RandomMat(3, 2, 4, 5), 1.f, 0, IntArray(-1))
|
|
|| test_reduction(RandomMat(3, 2, 4, 5), 2.f, 1, IntArray(-4, -2));
|
|
if (ret != 0)
|
|
return ret;
|
|
}
|
|
return 0;
|
|
}
|
|
|
|
static ncnn::Mat param_int_array(int size, int value)
|
|
{
|
|
ncnn::Mat m(size);
|
|
int* p = m;
|
|
for (int i = 0; i < size; i++)
|
|
p[i] = value;
|
|
|
|
return m;
|
|
}
|
|
|
|
#if NCNN_VALIDATION
|
|
static int test_reduction_load_param()
|
|
{
|
|
ncnn::ParamDict base;
|
|
base.set(5, 1); // fixbug0
|
|
base.set(3, param_int_array(1, 0));
|
|
if (test_layer_param(ncnn::LayerType::Reduction, base, 0)
|
|
|| test_layer_param(ncnn::LayerType::Reduction, base, 3, ncnn::Mat(0), 0))
|
|
return -1;
|
|
|
|
ncnn::Mat missing_data(0);
|
|
missing_data.w = 1;
|
|
|
|
return 0
|
|
|| test_layer_param(ncnn::LayerType::Reduction, base, 3, missing_data, -1)
|
|
|| test_layer_param(ncnn::LayerType::Reduction, base, 3, ncnn::Mat(1, (size_t)1u), -1)
|
|
|| test_layer_param(ncnn::LayerType::Reduction, base, 3, ncnn::Mat(1, 2), -1)
|
|
|| test_layer_param(ncnn::LayerType::Reduction, base, 3, 1.f, -1)
|
|
|| test_layer_param(ncnn::LayerType::Reduction, base, 3, param_int_array(5, 1), -1)
|
|
|| test_layer_param(ncnn::LayerType::Reduction, base, 3, param_int_array(1, INT_MIN), -1);
|
|
}
|
|
|
|
static int test_reduction_load_param_type()
|
|
{
|
|
ncnn::ParamDict base;
|
|
if (test_layer_param(ncnn::LayerType::Reduction, base, 0) != 0)
|
|
return -1;
|
|
|
|
for (int i = 0; i <= 10; i++)
|
|
{
|
|
if (test_layer_param(ncnn::LayerType::Reduction, base, 0, i, 0) != 0)
|
|
return -1;
|
|
}
|
|
|
|
const int invalid[] = {-1, 11, INT_MIN, INT_MAX};
|
|
for (int i = 0; i < 4; i++)
|
|
{
|
|
if (test_layer_param(ncnn::LayerType::Reduction, base, 0, invalid[i], -1) != 0)
|
|
return -1;
|
|
}
|
|
|
|
return 0;
|
|
}
|
|
|
|
static int test_reduction_load_param_text()
|
|
{
|
|
#if NCNN_STRING
|
|
const char* params[] = {"5=1 -23303=1,0", "5=1 -23303=1,0.0"};
|
|
for (int i = 0; i < 2; i++)
|
|
{
|
|
TestParamDict pd;
|
|
if (pd.load_param(params[i]) != 0 || pd.type(3) != 5 + i)
|
|
return -1;
|
|
|
|
if (test_layer_param(ncnn::LayerType::Reduction, pd, i == 0 ? 0 : -1) != 0)
|
|
return -1;
|
|
}
|
|
#endif
|
|
return 0;
|
|
}
|
|
#endif // NCNN_VALIDATION
|
|
|
|
static int test_reduction_depth_padding_boundaries()
|
|
{
|
|
// reducing depth isolates input channel padding from output depth indexing
|
|
ncnn::Mat a(1, 1, 2, 1);
|
|
a[0] = 1.f;
|
|
a[1] = 2.f;
|
|
|
|
// reducing height preserves two output depths with independent channel tail padding
|
|
ncnn::Mat b(1, 4, 2, 1);
|
|
for (int i = 0; i < 4; i++)
|
|
{
|
|
b[i] = 1.f;
|
|
b[4 + i] = 2.f;
|
|
}
|
|
|
|
op_type = 0;
|
|
return test_reduction(a, 1.f, 1, IntArray(1))
|
|
|| test_reduction(a, 1.f, 1, IntArray(-3))
|
|
|| test_reduction(b, 1.f, 1, IntArray(0, 2))
|
|
|| test_reduction(b, 1.f, 1, IntArray(-4, -2));
|
|
}
|
|
|
|
int main()
|
|
{
|
|
SRAND(7767517);
|
|
|
|
for (op_type = 0; op_type < OP_TYPE_MAX; op_type++)
|
|
{
|
|
int ret = 0
|
|
|| test_reduction_arithmetic()
|
|
|| test_reduction_vector_boundaries();
|
|
|
|
if (ret != 0)
|
|
return ret;
|
|
}
|
|
|
|
op_type = 0;
|
|
return test_reduction_geometry()
|
|
|| test_reduction_numeric_boundaries()
|
|
#if NCNN_VALIDATION
|
|
|| test_reduction_load_param()
|
|
|| test_reduction_load_param_type()
|
|
|| test_reduction_load_param_text()
|
|
#endif // NCNN_VALIDATION
|
|
|| test_reduction_depth_padding_boundaries();
|
|
}
|