mirror of
https://github.com/Tencent/ncnn.git
synced 2026-10-18 05:36:58 +08:00
nihui
41f031e045
* adjust tests * opt shader coverage op * update gpu_coverage.cpp atomicor coalesce
377 lines
16 KiB
C++
377 lines
16 KiB
C++
// Copyright 2019 Tencent
|
|
// SPDX-License-Identifier: BSD-3-Clause
|
|
|
|
#include "testutil.h"
|
|
|
|
#include "layer_type.h"
|
|
|
|
#include <limits.h>
|
|
|
|
static int test_deconvolutiondepthwise(int w, int h, int c, int outch, int kernel, int dilation, int stride, int pad, int bias, int group, int output_pad_right, int output_pad_bottom, int output_w, int output_h, int activation_type = 0, int flag = 0)
|
|
{
|
|
ncnn::Mat a = RandomMat(w, h, c);
|
|
|
|
if (output_w > 0 && output_h > 0 && pad != -233 && pad != -234)
|
|
{
|
|
pad = -233;
|
|
}
|
|
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, outch); // num_output
|
|
pd.set(1, kernel); // kernel_w
|
|
pd.set(2, dilation); // dilation_w
|
|
pd.set(3, stride); // stride_w
|
|
pd.set(4, pad); // pad_w
|
|
pd.set(5, bias); // bias_term
|
|
pd.set(6, outch / group * c / group * kernel * kernel * group);
|
|
pd.set(7, group);
|
|
|
|
ncnn::Mat activation_params(2);
|
|
activation_params[0] = activation_type == 2 ? 0.1f : -0.5f; // alpha
|
|
activation_params[1] = 0.25f; // beta
|
|
pd.set(9, activation_type);
|
|
pd.set(10, activation_params);
|
|
|
|
pd.set(18, output_pad_right);
|
|
pd.set(19, output_pad_bottom);
|
|
pd.set(20, output_w);
|
|
pd.set(21, output_h);
|
|
|
|
std::vector<ncnn::Mat> weights(2);
|
|
weights[0] = RandomMat(outch / group * c / group * kernel * kernel * group);
|
|
weights[1] = RandomMat(outch);
|
|
|
|
int ret = test_layer("DeconvolutionDepthWise", pd, weights, a, 0.001, flag);
|
|
if (ret != 0)
|
|
{
|
|
fprintf(stderr, "test_deconvolutiondepthwise failed w=%d h=%d c=%d outch=%d kernel=%d dilation=%d stride=%d pad=%d bias=%d group=%d act=%d actparams=[%f,%f] output_pad_right=%d output_pad_bottom=%d output_w=%d output_h=%d\n", w, h, c, outch, kernel, dilation, stride, pad, bias, group, activation_type, activation_params[0], activation_params[1], output_pad_right, output_pad_bottom, output_w, output_h);
|
|
}
|
|
|
|
return ret;
|
|
}
|
|
|
|
static int test_deconvolutiondepthwise_0()
|
|
{
|
|
// scalar and packed depthwise kernels cover all geometric boundaries
|
|
static const int kdsp[][4] = {
|
|
{1, 1, 1, 0}, {1, 1, 2, 0}, {2, 1, 1, 1}, {2, 1, 2, -233}, {3, 1, 1, 1}, {3, 1, 2, 1}, {3, 2, 1, 1}, {4, 1, 1, -233}, {4, 1, 2, -234}, {4, 2, 1, -234}, {5, 1, 1, 2}, {5, 1, 2, 2}, {5, 2, 2, 2}, {7, 1, 1, 3}, {7, 1, 2, 3}, {7, 2, 1, -233}
|
|
};
|
|
for (size_t i = 0; i < sizeof(kdsp) / sizeof(kdsp[0]); i++)
|
|
{
|
|
const int* k = kdsp[i];
|
|
if (test_deconvolutiondepthwise(15, 7, 1, 1, k[0], k[1], k[2], k[3], 1, 1, 0, 0, 0, 0)
|
|
|| test_deconvolutiondepthwise(15, 7, 8, 8, k[0], k[1], k[2], k[3], 0, 8, 0, 0, 0, 0))
|
|
return -1;
|
|
}
|
|
|
|
// channel remainders and crop axes use specialized and large generic representatives
|
|
static const int remainder_kdsp[][4] = {{1, 1, 1, 0}, {3, 1, 1, 1}, {7, 2, 1, -233}};
|
|
for (size_t i = 0; i < sizeof(remainder_kdsp) / sizeof(remainder_kdsp[0]); i++)
|
|
{
|
|
const int* k = remainder_kdsp[i];
|
|
if (test_deconvolutiondepthwise(15, 7, 2, 2, k[0], k[1], k[2], k[3], 1, 2, 1, 0, 0, 0)
|
|
|| test_deconvolutiondepthwise(15, 7, 3, 3, k[0], k[1], k[2], k[3], 0, 3, 0, 1, 0, 0)
|
|
|| test_deconvolutiondepthwise(15, 7, 4, 4, k[0], k[1], k[2], k[3], 0, 4, 2, 2, 0, 0)
|
|
|| test_deconvolutiondepthwise(15, 7, 7, 7, k[0], k[1], k[2], k[3], 1, 7, 2, 0, 0, 0)
|
|
|| test_deconvolutiondepthwise(15, 7, 15, 15, k[0], k[1], k[2], k[3], 1, 15, 3, 0, 7, 5))
|
|
return -1;
|
|
}
|
|
|
|
// group kernels retain all four arm specialized kernels and generic boundaries
|
|
static const int grouped_kdsp[][4] = {
|
|
{3, 1, 1, 1}, {3, 1, 2, 1}, {4, 1, 1, -233}, {4, 1, 2, -234}, {1, 1, 1, 0}, {3, 2, 1, 1}, {7, 2, 1, -233}
|
|
};
|
|
for (size_t i = 0; i < sizeof(grouped_kdsp) / sizeof(grouped_kdsp[0]); i++)
|
|
{
|
|
const int* k = grouped_kdsp[i];
|
|
if (test_deconvolutiondepthwise(15, 7, 2, 2, k[0], k[1], k[2], k[3], 0, 1, 1, 1, 7, 5)
|
|
|| test_deconvolutiondepthwise(15, 7, 4, 2, k[0], k[1], k[2], k[3], 1, 2, 0, 0, 7, 5)
|
|
|| test_deconvolutiondepthwise(15, 7, 8, 8, k[0], k[1], k[2], k[3], 0, 2, 0, 2, 7, 5)
|
|
|| test_deconvolutiondepthwise(15, 7, 12, 12, k[0], k[1], k[2], k[3], 0, 4, 3, 3, 0, 0)
|
|
|| test_deconvolutiondepthwise(15, 7, 16, 8, k[0], k[1], k[2], k[3], 0, 2, 0, 3, 0, 0))
|
|
return -1;
|
|
}
|
|
|
|
// activation types are explicit for every depthwise and group packing route
|
|
for (int activation = 0; activation < 5; activation++)
|
|
{
|
|
if (test_deconvolutiondepthwise(15, 7, 1, 1, 3, 1, 2, 1, 1, 1, 0, 0, 0, 0, activation)
|
|
|| test_deconvolutiondepthwise(15, 7, 4, 4, 3, 1, 2, 1, 0, 4, 2, 2, 0, 0, activation)
|
|
|| test_deconvolutiondepthwise(15, 7, 8, 8, 3, 1, 2, 1, 1, 8, 0, 0, 0, 0, activation)
|
|
|| test_deconvolutiondepthwise(15, 7, 2, 2, 3, 1, 2, 1, 0, 1, 1, 1, 7, 5, activation)
|
|
|| test_deconvolutiondepthwise(15, 7, 4, 2, 3, 1, 2, 1, 1, 2, 0, 0, 7, 5, activation)
|
|
|| test_deconvolutiondepthwise(15, 7, 8, 8, 3, 1, 2, 1, 0, 2, 0, 2, 7, 5, activation)
|
|
|| test_deconvolutiondepthwise(15, 7, 12, 12, 3, 1, 2, 1, 0, 4, 3, 3, 0, 0, activation)
|
|
|| test_deconvolutiondepthwise(15, 7, 16, 8, 3, 1, 2, 1, 0, 2, 0, 3, 0, 0, activation))
|
|
return -1;
|
|
}
|
|
|
|
// nonzero activation preserves each specialized scalar group's postprocessing
|
|
if (test_deconvolutiondepthwise(15, 7, 2, 2, 3, 1, 1, 1, 0, 1, 1, 1, 7, 5, 1)
|
|
|| test_deconvolutiondepthwise(15, 7, 2, 2, 3, 1, 2, 1, 0, 1, 1, 1, 7, 5, 1)
|
|
|| test_deconvolutiondepthwise(15, 7, 2, 2, 4, 1, 1, -233, 0, 1, 1, 1, 7, 5, 1)
|
|
|| test_deconvolutiondepthwise(15, 7, 2, 2, 4, 1, 2, -234, 0, 1, 1, 1, 7, 5, 1))
|
|
return -1;
|
|
|
|
return test_deconvolutiondepthwise(25, 13, 24, 24, 7, 2, 1, -233, 1, 24, 0, 0, 0, 0)
|
|
|| test_deconvolutiondepthwise(25, 13, 16, 8, 7, 2, 1, -233, 0, 2, 0, 3, 0, 0);
|
|
}
|
|
|
|
static int test_deconvolutiondepthwise_cpu_pack16()
|
|
{
|
|
// retain the original cpu-only pack16 group without adding gpu capability assumptions
|
|
static const int cases[][5] = {
|
|
{1, 1, 1, 0, 0}, {1, 1, 2, 0, 1}, {2, 1, 1, 1, 2}, {2, 1, 2, -233, 3}, {3, 1, 1, 1, 4}, {3, 1, 2, 1, 0}, {3, 2, 1, 1, 1}, {4, 1, 1, -233, 2}, {4, 1, 2, -234, 3}, {4, 2, 1, -234, 4}, {5, 1, 1, 2, 0}, {5, 1, 2, 2, 1}, {5, 2, 2, 2, 2}, {7, 1, 1, 3, 3}, {7, 1, 2, 3, 4}, {7, 2, 1, -233, 0}
|
|
};
|
|
for (size_t i = 0; i < sizeof(cases) / sizeof(cases[0]); i++)
|
|
{
|
|
const int* c = cases[i];
|
|
if (test_deconvolutiondepthwise(15, 7, 16, 16, c[0], c[1], c[2], c[3], 1, 16, 0, 0, 0, 0, c[4], TEST_LAYER_DISABLE_GPU_TESTING) != 0)
|
|
return -1;
|
|
}
|
|
return 0;
|
|
}
|
|
|
|
static int test_deconvolutiondepthwise_activation_params_text()
|
|
{
|
|
#if NCNN_STRING
|
|
const char* params[] = {"0=1 1=1 6=1 9=3 -23310=2,-1,2", "0=1 1=1 6=1 9=3 -23310=2,-1.0,2.0"};
|
|
for (int i = 0; i < 2; i++)
|
|
{
|
|
TestParamDict pd;
|
|
if (pd.load_param(params[i]) != 0 || pd.type(10) != 5 + i)
|
|
return -1;
|
|
|
|
std::vector<ncnn::Mat> weights(1);
|
|
weights[0].create(1);
|
|
weights[0][0] = 1.f;
|
|
|
|
ncnn::Mat a(2, 1, 1);
|
|
a[0] = -3.f;
|
|
a[1] = 3.f;
|
|
ncnn::Mat reference(2, 1, 1);
|
|
reference[0] = -1.f;
|
|
reference[1] = 2.f;
|
|
ncnn::Mat b;
|
|
int ret = test_layer_naive(ncnn::LayerType::DeconvolutionDepthWise, pd, weights, a, b, 0);
|
|
if (ret == 0)
|
|
ret = CompareMat(reference, b, 0.f);
|
|
if (ret != 0)
|
|
{
|
|
fprintf(stderr, "test_deconvolutiondepthwise_activation_params_text failed params=%s ret=%d\n", params[i], ret);
|
|
return ret;
|
|
}
|
|
|
|
const ncnn::Mat original = pd.get(10, ncnn::Mat());
|
|
const int* p = original;
|
|
if (i == 0 ? (p[0] != -1 || p[1] != 2) : (original[0] != -1.f || original[1] != 2.f))
|
|
{
|
|
fprintf(stderr, "test_deconvolutiondepthwise_activation_params_text modified params=%s\n", params[i]);
|
|
return -1;
|
|
}
|
|
}
|
|
#endif
|
|
return 0;
|
|
}
|
|
|
|
#if NCNN_VALIDATION
|
|
static int test_deconvolutiondepthwise_load_param_activation(const ncnn::ParamDict& base, int activation_type, const ncnn::Mat& activation_params, int expected_ret)
|
|
{
|
|
ncnn::ParamDict pd = base;
|
|
pd.set(9, activation_type);
|
|
pd.set(10, activation_params);
|
|
|
|
return test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, pd, expected_ret);
|
|
}
|
|
|
|
static int test_deconvolutiondepthwise_load_param()
|
|
{
|
|
ncnn::ParamDict base;
|
|
base.set(0, 8);
|
|
base.set(1, 3);
|
|
base.set(6, 216);
|
|
base.set(7, 2);
|
|
if (test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 0) != 0)
|
|
return -1;
|
|
|
|
ncnn::Mat params(2);
|
|
params[0] = 0.1f;
|
|
params[1] = 0.5f;
|
|
|
|
int ret = 0
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 0, ncnn::Mat(), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 0, ncnn::Mat(0), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 0, params, 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 0, params.range(0, 1), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 1, ncnn::Mat(), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 1, ncnn::Mat(0), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 1, params, 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 1, params.range(0, 1), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 2, ncnn::Mat(), -1)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 2, ncnn::Mat(0), -1)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 2, params, 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 2, params.range(0, 1), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 3, ncnn::Mat(), -1)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 3, ncnn::Mat(0), -1)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 3, params, 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 3, params.range(0, 1), -1)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 4, ncnn::Mat(), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 4, ncnn::Mat(0), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 4, params, 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 4, params.range(0, 1), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 5, ncnn::Mat(), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 5, ncnn::Mat(0), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 5, params, 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 5, params.range(0, 1), 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 6, ncnn::Mat(), -1)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 6, ncnn::Mat(0), -1)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 6, params, 0)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 6, params.range(0, 1), -1)
|
|
|| test_deconvolutiondepthwise_load_param_activation(base, 7, ncnn::Mat(), -1);
|
|
if (ret != 0)
|
|
return ret;
|
|
|
|
if (test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 10, 1, -1)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 10, 1.f, -1))
|
|
return -1;
|
|
|
|
ncnn::Mat missing_data(0);
|
|
missing_data.w = 1;
|
|
|
|
const ncnn::Mat bad[] = {ncnn::Mat(2, (size_t)1u), ncnn::Mat(2, (size_t)2u), ncnn::Mat(2, 2), ncnn::Mat(2, (size_t)16u, 4), missing_data};
|
|
for (int i = 0; i < 5; i++)
|
|
{
|
|
if (test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 10, bad[i], -1) != 0)
|
|
return -1;
|
|
}
|
|
|
|
ret = 0
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 0, 0, -1)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 1, 0, -1)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 2, 0, -1)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 3, 0, -1)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 1, INT_MAX, -1)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 6, 217, -1);
|
|
if (ret != 0)
|
|
return ret;
|
|
|
|
const int groups[] = {0, -1, -8, INT_MIN, 3};
|
|
for (int i = 0; i < 5; i++)
|
|
{
|
|
if (test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 7, groups[i], -1) != 0)
|
|
return -1;
|
|
}
|
|
|
|
{
|
|
ncnn::ParamDict pd = base;
|
|
pd.set(0, INT_MIN);
|
|
pd.set(7, -1); // avoid evaluating INT_MIN % -1
|
|
if (test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, pd, -1) != 0)
|
|
return -1;
|
|
}
|
|
|
|
return 0;
|
|
}
|
|
|
|
static int test_deconvolutiondepthwise_load_param_dynamic()
|
|
{
|
|
ncnn::ParamDict base;
|
|
base.set(0, 5);
|
|
base.set(1, 1);
|
|
base.set(6, 10);
|
|
base.set(7, 2);
|
|
if (test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, -1) != 0)
|
|
return -1;
|
|
|
|
ncnn::ParamDict dynamic = base;
|
|
dynamic.set(28, 1);
|
|
|
|
const int num_outputs[] = {0, -1, 5, INT_MIN, INT_MAX};
|
|
for (int i = 0; i < 5; i++)
|
|
{
|
|
if (test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, dynamic, 0, num_outputs[i], 0) != 0)
|
|
return -1;
|
|
}
|
|
|
|
return 0
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, base, 0, 2, 0)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, dynamic, 7, 0, -1)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, dynamic, 7, -1, -1)
|
|
|| test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, dynamic, 7, INT_MIN, -1);
|
|
}
|
|
|
|
static int test_deconvolutiondepthwise_load_param_text()
|
|
{
|
|
#if NCNN_STRING
|
|
const char* params[] = {"0=1 1=1 6=1 9=3 -23310=2,-1,2", "0=1 1=1 6=1 9=3 -23310=2,-1.0,2.0"};
|
|
for (int i = 0; i < 2; i++)
|
|
{
|
|
TestParamDict pd;
|
|
if (pd.load_param(params[i]) != 0 || pd.type(10) != 5 + i)
|
|
return -1;
|
|
|
|
if (test_layer_param(ncnn::LayerType::DeconvolutionDepthWise, pd, 0) != 0)
|
|
return -1;
|
|
}
|
|
#endif
|
|
return 0;
|
|
}
|
|
#endif // NCNN_VALIDATION
|
|
|
|
static int test_deconvolutiondepthwise_relu_scalar_tail()
|
|
{
|
|
// grouped scalar output has nine elements and a deterministic negative fp16 tail
|
|
ncnn::Mat a(1, 1, 2);
|
|
a.fill(-1.f);
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, 2);
|
|
pd.set(1, 3);
|
|
pd.set(2, 1);
|
|
pd.set(3, 2);
|
|
pd.set(4, 0);
|
|
pd.set(5, 0);
|
|
pd.set(6, 36);
|
|
pd.set(7, 1);
|
|
pd.set(9, 1);
|
|
ncnn::Mat activation_params(2);
|
|
activation_params[0] = 0.1f;
|
|
activation_params[1] = 0.25f;
|
|
pd.set(10, activation_params);
|
|
std::vector<ncnn::Mat> weights(1);
|
|
weights[0].create(36);
|
|
weights[0].fill(1.f);
|
|
return test_layer("DeconvolutionDepthWise", pd, weights, a, 0.001);
|
|
}
|
|
|
|
static int test_deconvolutiondepthwise_group_pack_bias()
|
|
{
|
|
// both per-group packing directions retain their opposite bias boundary
|
|
return 0
|
|
|| test_deconvolutiondepthwise(1, 1, 6, 8, 1, 1, 1, 0, 1, 2, 0, 0, 0, 0, 3)
|
|
|| test_deconvolutiondepthwise(1, 1, 8, 6, 1, 1, 1, 0, 0, 2, 0, 0, 0, 0, 3);
|
|
}
|
|
|
|
int main()
|
|
{
|
|
SRAND(7767517);
|
|
|
|
return 0
|
|
|| test_deconvolutiondepthwise_0()
|
|
|| test_deconvolutiondepthwise_cpu_pack16()
|
|
|| test_deconvolutiondepthwise_activation_params_text()
|
|
#if NCNN_VALIDATION
|
|
|| test_deconvolutiondepthwise_load_param()
|
|
|| test_deconvolutiondepthwise_load_param_text()
|
|
|| test_deconvolutiondepthwise_load_param_dynamic()
|
|
#endif // NCNN_VALIDATION \
|
|
// the grouped k2 kernel produces an internal gemm M8 packing tier
|
|
|| test_deconvolutiondepthwise(15, 7, 2, 2, 2, 1, 1, 1, 0, 1, 1, 1, 7, 5)
|
|
|| test_deconvolutiondepthwise_relu_scalar_tail()
|
|
// each group exercises scalar to packed and packed to scalar kernels
|
|
|| test_deconvolutiondepthwise(7, 5, 6, 8, 3, 1, 1, 1, 0, 2, 0, 0, 0, 0, 3)
|
|
|| test_deconvolutiondepthwise(7, 5, 8, 6, 3, 1, 2, 1, 1, 2, 1, 0, 0, 0, 3)
|
|
// same lower explicitly crops an odd border in both axes
|
|
|| test_deconvolutiondepthwise(2, 2, 1, 1, 4, 1, 2, -234, 0, 1, 0, 0, 3, 3, 0)
|
|
|| test_deconvolutiondepthwise_group_pack_bias();
|
|
}
|