Skip to content

O
opencv
Commit f6c57388 authored by Alexander Alekhin's avatar Alexander Alekhin
Browse files
Options

Merge remote-tracking branch 'upstream/3.4' into merge-3.4

parents f663e8f9 054c7962
Hide whitespace changes
Inline Side-by-side
Showing with 638 additions and 228 deletions
modules/dnn/perf/perf_convolution3d.cpp 0 → 100644
View file @ f6c57388
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#include "perf_precomp.hpp"
#include <opencv2/dnn/shape_utils.hpp>
namespace opencv_test {
struct Conv3DParam_t {
int kernel[3];
struct BlobShape { int dims[5]; } shapeIn;
int outCN;
int groups;
int stride[3];
int dilation[3];
int pad[6];
const char* padMode;
bool hasBias;
double declared_flops;
};
// Details: #12142
static const Conv3DParam_t testConvolution3DConfigs[] = {
{{3, 3, 3}, {{1, 6, 10, 38, 50}}, 6, 1, {1, 1, 1}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "VALID", true, 26956800.},
{{3, 3, 3}, {{1, 2, 19, 19, 19}}, 2, 2, {2, 2, 2}, {1, 1, 1}, {1, 1, 1, 1, 1, 1}, "", true, 218000.},
{{3, 3, 3}, {{1, 2, 25, 19, 19}}, 2, 2, {1, 2, 2}, {1, 1, 1}, {2, 2, 2, 2, 2, 2}, "SAME", false, 545000.},
{{3, 3, 3}, {{1, 11, 9, 150, 200}}, 11, 1, {1, 1, 1}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "VALID", true, 1342562760.},
{{3, 3, 3}, {{1, 10, 98, 10, 10}}, 10, 1, {1, 1, 1}, {1, 1, 1}, {1, 0, 1, 1, 0,1}, "SAME", false, 53018000.},
{{5, 5, 5}, {{1, 6, 19, 19, 19}}, 6, 2, {1, 1, 1}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "", false, 30395250.},
{{5, 5, 5}, {{1, 4, 50, 19, 19}}, 4, 1, {2, 2, 2}, {1, 1, 1}, {1, 1, 1, 1, 1, 1}, "VALID", false, 5893888.},
{{5, 5, 5}, {{1, 3, 75, 75, 100}}, 3, 1, {1, 1, 1}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "SAME", true, 1267312500.},
{{5, 5, 5}, {{1, 2, 21, 75, 100}}, 2, 1, {1, 1, 1}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "", true, 116103744.},
{{5, 5, 5}, {{1, 4, 40, 75, 75}}, 4, 1, {2, 2, 2}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "", false, 93405312.},
{{7, 7, 7}, {{1, 6, 15, 19, 19}}, 6, 1, {2, 1, 1}, {1, 1, 1}, {3, 3, 3, 3, 3, 3}, "SAME", true, 71339376.},
{{7, 7, 7}, {{1, 2, 38, 38, 38}}, 2, 1, {1, 2, 1}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "", false, 44990464.},
{{1, 1, 1}, {{1, 4, 9, 10, 10}}, 4, 1, {1, 1, 2}, {1, 1, 1}, {1, 1, 1, 1, 1, 1}, "VALID", false, 16200.},
{{3, 1, 4}, {{1, 14, 5, 10, 10}}, 14, 1, {1, 1, 1}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "SAME", false, 2359000.},
{{1, 1, 1}, {{1, 8, 1, 10, 10}}, 8, 8, {1, 1, 1}, {1, 1, 1}, {1, 1, 1, 1, 1, 1}, "", true, 58752.},
{{3, 4, 2}, {{1, 4, 8, 10, 10}}, 4, 4, {1, 2, 1}, {1, 1, 1}, {0, 0, 0, 0, 0, 0}, "", true, 166752.}
};
struct Conv3DParamID
{
enum {
CONV_0 = 0,
CONV_100 = 16,
CONV_LAST = sizeof(testConvolution3DConfigs) / sizeof(testConvolution3DConfigs[0])
};
int val_; \
Conv3DParamID(int val = 0) : val_(val) {}
operator int() const { return val_; }
static ::testing::internal::ParamGenerator<Conv3DParamID> all()
{
#if 0
enum { NUM = (int)CONV_LAST };
#else
enum { NUM = (int)CONV_100 };
#endif
Conv3DParamID v_[NUM]; for (int i = 0; i < NUM; ++i) { v_[i] = Conv3DParamID(i); } // reduce generated code size
return ::testing::ValuesIn(v_, v_ + NUM);
}
}; \
static inline void PrintTo(const Conv3DParamID& v, std::ostream* os)
{
CV_Assert((int)v >= 0); CV_Assert((int)v < Conv3DParamID::CONV_LAST);
const Conv3DParam_t& p = testConvolution3DConfigs[(int)v];
*os << "GFLOPS=" << cv::format("%.3f", p.declared_flops * 1e-9)
<< ", K=[" << p.kernel[0] << " x " << p.kernel[1] << " x " << p.kernel[2] << "]"
<< ", IN={" << p.shapeIn.dims[0] << ", " << p.shapeIn.dims[1] << ", " << p.shapeIn.dims[2] << ", " << p.shapeIn.dims[3] << ", " << p.shapeIn.dims[4] << "}"
<< ", OCN=" << p.outCN;
if (p.groups > 1)
*os << ", G=" << p.groups;
if (p.stride[0] * p.stride[1] * p.stride[2] != 1)
*os << ", S=[" << p.stride[0] << " x " << p.stride[1] << " x " << p.stride[2] << "]";
if (p.dilation[0] * p.dilation[1] * p.dilation[2] != 1)
*os << ", D=[" << p.dilation[0] << " x " << p.dilation[1] << " x " << p.dilation[2] << "]";
if (p.pad[0] != 0 && p.pad[1] != 0 && p.pad[2] != 0 &&
p.pad[3] != 0 && p.pad[4] != 0 && p.pad[5] != 0)
*os << ", P=(" << p.pad[0] << ", " << p.pad[3] << ") x ("
<< p.pad[1] << ", " << p.pad[4] << ") x ("
<< p.pad[2] << ", " << p.pad[5] << ")";
if (!((std::string)p.padMode).empty())
*os << ", PM=" << ((std::string)p.padMode);
if (p.hasBias)
*os << ", BIAS";
}
typedef tuple<Conv3DParamID, tuple<Backend, Target> > Conv3DTestParam_t;
typedef TestBaseWithParam<Conv3DTestParam_t> Conv3D;
PERF_TEST_P_(Conv3D, conv3d)
{
int test_id = (int)get<0>(GetParam());
ASSERT_GE(test_id, 0); ASSERT_LT(test_id, Conv3DParamID::CONV_LAST);
const Conv3DParam_t& params = testConvolution3DConfigs[test_id];
double declared_flops = params.declared_flops;
DictValue kernel = DictValue::arrayInt(&params.kernel[0], 3);
DictValue stride = DictValue::arrayInt(&params.stride[0], 3);
DictValue pad = DictValue::arrayInt(&params.pad[0], 6);
DictValue dilation = DictValue::arrayInt(&params.dilation[0], 3);
MatShape inputShape = MatShape(params.shapeIn.dims, params.shapeIn.dims + 5);
int outChannels = params.outCN;
int groups = params.groups;
std::string padMode(params.padMode);
bool hasBias = params.hasBias;
Backend backendId = get<0>(get<1>(GetParam()));
Target targetId = get<1>(get<1>(GetParam()));
if (targetId != DNN_TARGET_CPU)
throw SkipTestException("Only CPU is supported");
int inChannels = inputShape[1];
int sz[] = {outChannels, inChannels / groups, params.kernel[0], params.kernel[1], params.kernel[2]};
Mat weights(5, &sz[0], CV_32F);
randu(weights, -1.0f, 1.0f);
LayerParams lp;
lp.set("kernel_size", kernel);
lp.set("pad", pad);
if (!padMode.empty())
lp.set("pad_mode", padMode);
lp.set("stride", stride);
lp.set("dilation", dilation);
lp.set("num_output", outChannels);
lp.set("group", groups);
lp.set("bias_term", hasBias);
lp.type = "Convolution";
lp.name = "testLayer";
lp.blobs.push_back(weights);
if (hasBias)
{
Mat bias(1, outChannels, CV_32F);
randu(bias, -1.0f, 1.0f);
lp.blobs.push_back(bias);
}
int inpSz[] = {1, inChannels, inputShape[2], inputShape[3], inputShape[4]};
Mat input(5, &inpSz[0], CV_32F);
randu(input, -1.0f, 1.0f);
Net net;
net.addLayerToPrev(lp.name, lp.type, lp);
net.setInput(input);
net.setPreferableBackend(backendId);
net.setPreferableTarget(targetId);
Mat output = net.forward();
MatShape netInputShape = shape(input);
size_t weightsMemory = 0, blobsMemory = 0;
net.getMemoryConsumption(netInputShape, weightsMemory, blobsMemory);
int64 flops = net.getFLOPS(netInputShape);
CV_Assert(flops > 0);
std::cout
<< "IN=" << divUp(input.total() * input.elemSize(), 1u<<10) << " Kb " << netInputShape
<< " OUT=" << divUp(output.total() * output.elemSize(), 1u<<10) << " Kb " << shape(output)
<< " Weights(parameters): " << divUp(weightsMemory, 1u<<10) << " Kb"
<< " MFLOPS=" << flops * 1e-6 << std::endl;
TEST_CYCLE()
{
Mat res = net.forward();
}
EXPECT_NEAR(flops, declared_flops, declared_flops * 1e-6);
SANITY_CHECK_NOTHING();
}
INSTANTIATE_TEST_CASE_P(/**/, Conv3D, Combine(
Conv3DParamID::all(),
dnnBackendsAndTargets(false, false) // defined in ../test/test_common.hpp
));
} // namespace
modules/dnn/src/layers/convolution_layer.cpp
View file @ f6c57388
...@@ -48,6 +48,7 @@ ...@@ -48,6 +48,7 @@
#include "opencv2/core/hal/hal.hpp" #include "opencv2/core/hal/hal.hpp"
#include "opencv2/core/hal/intrin.hpp" #include "opencv2/core/hal/intrin.hpp"
#include <iostream> #include <iostream>
#include <numeric>
#ifdef HAVE_OPENCL #ifdef HAVE_OPENCL
#include "opencl_kernels_dnn.hpp" #include "opencl_kernels_dnn.hpp"
...@@ -67,7 +68,7 @@ public: ...@@ -67,7 +68,7 @@ public:
BaseConvolutionLayerImpl(const LayerParams &params) BaseConvolutionLayerImpl(const LayerParams &params)
{ {
setParamsFrom(params); setParamsFrom(params);
getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations, padMode); getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations, padMode, adjust_pads);
numOutput = params.get<int>("num_output"); numOutput = params.get<int>("num_output");
int ngroups = params.get<int>("group", 1); int ngroups = params.get<int>("group", 1);
...@@ -83,14 +84,14 @@ public: ...@@ -83,14 +84,14 @@ public:
pad = Size(pads_begin[1], pads_begin[0]); pad = Size(pads_begin[1], pads_begin[0]);
dilation = Size(dilations[1], dilations[0]); dilation = Size(dilations[1], dilations[0]);
adjust_pads.push_back(params.get<int>("adj_h", 0));
adjust_pads.push_back(params.get<int>("adj_w", 0));
adjustPad.height = adjust_pads[0]; adjustPad.height = adjust_pads[0];
adjustPad.width = adjust_pads[1]; adjustPad.width = adjust_pads[1];
CV_Assert(adjustPad.width < stride.width &&
adjustPad.height < stride.height);
} }
for (int i = 0; i < adjust_pads.size(); i++) {
CV_Assert(adjust_pads[i] < strides[i]);
}
fusedWeights = false; fusedWeights = false;
fusedBias = false; fusedBias = false;
} }
...@@ -258,11 +259,14 @@ public: ...@@ -258,11 +259,14 @@ public:
else else
#endif #endif
{ {
if (kernel_size.size() != 2) if (kernel_size.size() == 3)
return (preferableTarget == DNN_TARGET_CPU && backendId == DNN_BACKEND_OPENCV);
else if (kernel_size.size() == 2)
return backendId == DNN_BACKEND_OPENCV ||
backendId == DNN_BACKEND_HALIDE ||
(backendId == DNN_BACKEND_VKCOM && haveVulkan());
else
return false; return false;
return backendId == DNN_BACKEND_OPENCV ||
backendId == DNN_BACKEND_HALIDE ||
(backendId == DNN_BACKEND_VKCOM && haveVulkan());
} }
} }
...@@ -604,8 +608,8 @@ public: ...@@ -604,8 +608,8 @@ public:
const Mat* input_; const Mat* input_;
const Mat* weights_; const Mat* weights_;
Mat* output_; Mat* output_;
int outShape[4]; int outShape[4]; // used only for conv2d
Size kernel_, pad_, stride_, dilation_; std::vector<size_t> kernel_size, pads_begin, pads_end, strides, dilations;
int ngroups_, nstripes_; int ngroups_, nstripes_;
std::vector<int> ofstab_; std::vector<int> ofstab_;
const std::vector<float>* biasvec_; const std::vector<float>* biasvec_;
...@@ -624,14 +628,18 @@ public: ...@@ -624,14 +628,18 @@ public:
static void run( const Mat& input, Mat& output, const Mat& weights, static void run( const Mat& input, Mat& output, const Mat& weights,
const std::vector<float>& biasvec, const std::vector<float>& biasvec,
const std::vector<float>& reluslope, const std::vector<float>& reluslope,
Size kernel, Size pad, Size stride, Size dilation, const std::vector<size_t>& kernel_size, const std::vector<size_t>& strides,
const std::vector<size_t>& pads_begin, const std::vector<size_t>& pads_end,
const std::vector<size_t>& dilations,
const ActivationLayer* activ, int ngroups, int nstripes ) const ActivationLayer* activ, int ngroups, int nstripes )
{ {
size_t karea = std::accumulate(kernel_size.begin(), kernel_size.end(),
1, std::multiplies<size_t>());
CV_Assert_N( CV_Assert_N(
input.dims == 4 && output.dims == 4, (input.dims == 4 || input.dims == 5) && (input.dims == output.dims),
input.size[0] == output.size[0], input.size[0] == output.size[0],
weights.rows == output.size[1], weights.rows == output.size[1],
weights.cols == (input.size[1]/ngroups)*kernel.width*kernel.height, weights.cols == (input.size[1]/ngroups)*karea,
input.type() == output.type(), input.type() == output.type(),
input.type() == weights.type(), input.type() == weights.type(),
input.type() == CV_32FC1, input.type() == CV_32FC1,
...@@ -645,26 +653,58 @@ public: ...@@ -645,26 +653,58 @@ public:
p.output_ = &output; p.output_ = &output;
for( int i = 0; i < 4; i++ ) p.outShape[i] = output.size[i]; for( int i = 0; i < 4; i++ ) p.outShape[i] = output.size[i];
p.outShape[1] /= ngroups; p.outShape[1] /= ngroups;
p.kernel_ = kernel; p.pad_ = pad; p.stride_ = stride; p.dilation_ = dilation;
p.kernel_size = kernel_size; p.strides = strides; p.dilations = dilations;
p.pads_begin = pads_begin; p.pads_end = pads_end;
p.ngroups_ = ngroups; p.ngroups_ = ngroups;
p.nstripes_ = nstripes; p.nstripes_ = nstripes;
int inpCnAll = input.size[1], width = input.size[3], height = input.size[2]; int inpCnAll = input.size[1];
int depth = (input.dims == 5) ? input.size[2] : 1;
int width = input.size[input.dims - 1];
int height = input.size[input.dims - 2];
int inpCn = inpCnAll / ngroups; int inpCn = inpCnAll / ngroups;
p.is1x1_ = kernel == Size(1,1) && pad == Size(0, 0);
p.useAVX = checkHardwareSupport(CPU_AVX); bool isConv2D = kernel_size.size() == 2;
p.useAVX2 = checkHardwareSupport(CPU_AVX2);
p.useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX; p.is1x1_ = isConv2D && kernel_size[0] == 1 && kernel_size[1] == 1 &&
pads_begin[0] == 0 && pads_begin[1] == 0;
p.useAVX = checkHardwareSupport(CPU_AVX) && isConv2D;
p.useAVX2 = checkHardwareSupport(CPU_AVX2) && isConv2D;
p.useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX && isConv2D;
int ncn = std::min(inpCn, (int)BLK_SIZE_CN); int ncn = std::min(inpCn, (int)BLK_SIZE_CN);
p.ofstab_.resize(kernel.width*kernel.height*ncn);
int kernel_d = !isConv2D? kernel_size[0] : 1;
int kernel_h = kernel_size[kernel_size.size() - 2];
int kernel_w = kernel_size.back();
int dil_d = !isConv2D? dilations[0] : 1;
int dil_h = dilations[dilations.size() - 2];
int dil_w = dilations.back();
p.ofstab_.resize(karea * ncn);
int* ofstab = &p.ofstab_[0]; int* ofstab = &p.ofstab_[0];
for( int k = 0; k < ncn; k++ ) if (isConv2D)
for( int k_r = 0; k_r < kernel.height; k_r++ ) {
for( int k_c = 0; k_c < kernel.width; k_c++ ) for( int k = 0; k < ncn; k++ )
ofstab[(k*kernel.height + k_r)*kernel.width + k_c] = for( int k_r = 0; k_r < kernel_h; k_r++ )
(k*height + k_r*dilation.height)*width + k_c*dilation.width; for( int k_c = 0; k_c < kernel_w; k_c++ )
ofstab[(k*kernel_h + k_r)*kernel_w + k_c] =
(k*height + k_r*dil_h)*width + k_c*dil_w;
}
else
{
for( int k = 0; k < ncn; k++ )
for (int k_d = 0; k_d < kernel_d; k_d++)
for( int k_r = 0; k_r < kernel_h; k_r++ )
for( int k_c = 0; k_c < kernel_w; k_c++ )
ofstab[(k*kernel_d*kernel_h + k_d*kernel_h + k_r)*kernel_w + k_c] =
(k*depth*height + k_d*dil_d*height + k_r*dil_h)*width + k_c*dil_w;
}
p.biasvec_ = &biasvec; p.biasvec_ = &biasvec;
p.reluslope_ = &reluslope; p.reluslope_ = &reluslope;
...@@ -677,17 +717,39 @@ public: ...@@ -677,17 +717,39 @@ public:
{ {
const int valign = ConvolutionLayerImpl::VEC_ALIGN; const int valign = ConvolutionLayerImpl::VEC_ALIGN;
int ngroups = ngroups_, batchSize = input_->size[0]*ngroups; int ngroups = ngroups_, batchSize = input_->size[0]*ngroups;
int outW = output_->size[3], outH = output_->size[2], outCn = output_->size[1]/ngroups; bool isConv2D = input_->dims == 4;
int width = input_->size[3], height = input_->size[2], inpCn = input_->size[1]/ngroups;
int outW = output_->size[output_->dims - 1];
int outH = output_->size[output_->dims - 2];
int outCn = output_->size[1]/ngroups;
int depth = !isConv2D? input_->size[2] : 1;
int height = input_->size[input_->dims - 2];
int width = input_->size[input_->dims - 1];
int inpCn = input_->size[1]/ngroups;
const int nstripes = nstripes_; const int nstripes = nstripes_;
int kernel_w = kernel_.width, kernel_h = kernel_.height;
int pad_w = pad_.width, pad_h = pad_.height; int kernel_d = !isConv2D? kernel_size[0] : 1;
int stride_w = stride_.width, stride_h = stride_.height; int kernel_h = kernel_size[kernel_size.size() - 2];
int dilation_w = dilation_.width, dilation_h = dilation_.height; int kernel_w = kernel_size.back();
int karea = kernel_w*kernel_h; int karea = kernel_w*kernel_h*kernel_d;
int i, j, k;
size_t inpPlaneSize = width*height; int pad_d = !isConv2D? pads_begin[0] : 0;
size_t outPlaneSize = outW*outH; int pad_t = pads_begin[pads_begin.size() - 2];
int pad_l = pads_begin.back();
int stride_d = !isConv2D? strides[0] : 0;
int stride_h = strides[strides.size() - 2];
int stride_w = strides.back();
int dilation_d = !isConv2D? dilations[0] : 1;
int dilation_h = dilations[dilations.size() - 2];
int dilation_w = dilations.back();
int i, j, k, d;
size_t inpPlaneSize = input_->total(2);
size_t outPlaneSize = output_->total(2);
bool is1x1 = is1x1_; bool is1x1 = is1x1_;
int stripesPerSample; int stripesPerSample;
...@@ -756,72 +818,125 @@ public: ...@@ -756,72 +818,125 @@ public:
for( int ofs0 = stripeStart; ofs0 < stripeEnd; ofs0 += BLK_SIZE ) for( int ofs0 = stripeStart; ofs0 < stripeEnd; ofs0 += BLK_SIZE )
{ {
int ofs, ofs1 = std::min(ofs0 + BLK_SIZE, stripeEnd); int ofs, ofs1 = std::min(ofs0 + BLK_SIZE, stripeEnd);
int out_i = ofs0 / outW;
int out_j = ofs0 - out_i * outW; int out_d = ofs0 / (outH * outW);
int out_i = (ofs0 - out_d * outH * outW) / outW;
int out_j = ofs0 % outW;
// do im2row for a part of input tensor // do im2row for a part of input tensor
float* rowbuf = rowbuf0; float* rowbuf = rowbuf0;
for( ofs = ofs0; ofs < ofs1; out_j = 0, ++out_i )
if (isConv2D)
{ {
int delta = std::min(ofs1 - ofs, outW - out_j); for( ofs = ofs0; ofs < ofs1; out_j = 0, ++out_i )
int out_j1 = out_j + delta;
int in_i = out_i * stride_h - pad_h;
int in_j = out_j * stride_w - pad_w;
const float* imgptr = data_inp0 + (cn0*height + in_i)*width + in_j;
ofs += delta;
// do im2row for a part of input tensor
if( is1x1 )
{ {
for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w ) int delta = std::min(ofs1 - ofs, outW - out_j);
int out_j1 = out_j + delta;
int in_i = out_i * stride_h - pad_t;
int in_j = out_j * stride_w - pad_l;
const float* imgptr = data_inp0 + (cn0*height + in_i)*width + in_j;
ofs += delta;
// do im2row for a part of input tensor
if( is1x1 )
{ {
for( k = 0; k < vsz; k++ ) for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w )
rowbuf[k] = imgptr[k*inpPlaneSize]; {
for( k = 0; k < vsz; k++ )
rowbuf[k] = imgptr[k*inpPlaneSize];
}
}
else
{
bool ok_i = 0 <= in_i && in_i < height - (kernel_h-1)*dilation_h;
int i0 = std::max(0, (-in_i + dilation_h-1)/dilation_h);
int i1 = std::min(kernel_h, (height - in_i + dilation_h-1)/dilation_h);
for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w, in_j += stride_w )
{
// this condition should be true for most of the tensor elements, i.e.
// most of the time the kernel aperture is inside the tensor X-Y plane.
if( ok_i && out_j + 2 <= out_j1 && 0 <= in_j && in_j + stride_w*2 <= width - (kernel_w-1)*dilation_w )
{
for( k = 0; k < vsz; k++ )
{
int k1 = ofstab[k];
float v0 = imgptr[k1];
float v1 = imgptr[k1 + stride_w];
rowbuf[k] = v0;
rowbuf[k+vsz_a] = v1;
}
out_j++;
rowbuf += vsz_a;
imgptr += stride_w;
in_j += stride_w;
}
else
{
int j0 = std::max(0, (-in_j + dilation_w-1)/dilation_w);
int j1 = std::min(kernel_w, (width - in_j + dilation_w-1)/dilation_w);
// here some non-continuous sub-row of the row will not be
// filled from the tensor; we need to make sure that the uncovered
// elements are explicitly set to 0's. the easiest way is to
// set all the elements to 0's before the loop.
memset(rowbuf, 0, vsz*sizeof(rowbuf[0]));
for( k = 0; k < ncn; k++ )
{
for( i = i0; i < i1; i++ )
{
for( j = j0; j < j1; j++ )
{
int imgofs = k*(width*height) + i*(dilation_h*width) + j*dilation_w;
rowbuf[(k*kernel_h + i)*kernel_w + j] = imgptr[imgofs];
}
}
}
}
}
} }
} }
else }
else
{
for( ofs = ofs0; ofs < ofs1; out_d += (out_i + 1) / outH, out_i = (out_i + 1) % outH, out_j = 0 )
{ {
bool ok_i = 0 <= in_i && in_i < height - (kernel_h-1)*dilation_h; int delta = std::min(ofs1 - ofs, outW - out_j);
int out_j1 = out_j + delta;
int in_d = out_d * stride_d - pad_d;
int in_i = out_i * stride_h - pad_t;
int in_j = out_j * stride_w - pad_l;
const float* imgptr = data_inp0 + (cn0*depth*height + in_d*height + in_i)*width + in_j;
ofs += delta;
int d0 = std::max(0, (-in_d + dilation_d - 1) / dilation_d);
int d1 = std::min(kernel_d, (depth - in_d + dilation_d - 1) / dilation_d);
int i0 = std::max(0, (-in_i + dilation_h-1)/dilation_h); int i0 = std::max(0, (-in_i + dilation_h-1)/dilation_h);
int i1 = std::min(kernel_h, (height - in_i + dilation_h-1)/dilation_h); int i1 = std::min(kernel_h, (height - in_i + dilation_h-1)/dilation_h);
for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w, in_j += stride_w ) for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w, in_j += stride_w )
{ {
// this condition should be true for most of the tensor elements, i.e. int j0 = std::max(0, (-in_j + dilation_w-1)/dilation_w);
// most of the time the kernel aperture is inside the tensor X-Y plane. int j1 = std::min(kernel_w, (width - in_j + dilation_w-1)/dilation_w);
if( ok_i && out_j + 2 <= out_j1 && 0 <= in_j && in_j + stride_w*2 <= width - (kernel_w-1)*dilation_w )
// here some non-continuous sub-row of the row will not be
// filled from the tensor; we need to make sure that the uncovered
// elements are explicitly set to 0's. the easiest way is to
// set all the elements to 0's before the loop.
memset(rowbuf, 0, vsz*sizeof(rowbuf[0]));
for( k = 0; k < ncn; k++ )
{ {
for( k = 0; k < vsz; k++ ) for ( d = d0; d < d1; d++)
{
int k1 = ofstab[k];
float v0 = imgptr[k1];
float v1 = imgptr[k1 + stride_w];
rowbuf[k] = v0;
rowbuf[k+vsz_a] = v1;
}
out_j++;
rowbuf += vsz_a;
imgptr += stride_w;
in_j += stride_w;
}
else
{
int j0 = std::max(0, (-in_j + dilation_w-1)/dilation_w);
int j1 = std::min(kernel_w, (width - in_j + dilation_w-1)/dilation_w);
// here some non-continuous sub-row of the row will not be
// filled from the tensor; we need to make sure that the uncovered
// elements are explicitly set to 0's. the easiest way is to
// set all the elements to 0's before the loop.
memset(rowbuf, 0, vsz*sizeof(rowbuf[0]));
for( k = 0; k < ncn; k++ )
{ {
for( i = i0; i < i1; i++ ) for( i = i0; i < i1; i++ )
{ {
for( j = j0; j < j1; j++ ) for( j = j0; j < j1; j++ )
{ {
int imgofs = k*(width*height) + i*(dilation_h*width) + j*dilation_w; int imgofs = k*(depth*width*height) + d*dilation_d*width*height + i*(dilation_h*width) + j*dilation_w;
rowbuf[(k*kernel_h + i)*kernel_w + j] = imgptr[imgofs]; rowbuf[(k*kernel_d*kernel_h + d*kernel_h + i)*kernel_w + j] = imgptr[imgofs];
} }
} }
} }
...@@ -1131,10 +1246,6 @@ public: ...@@ -1131,10 +1246,6 @@ public:
CV_Assert_N(inputs.size() == (size_t)1, inputs[0].size[1] % blobs[0].size[1] == 0, CV_Assert_N(inputs.size() == (size_t)1, inputs[0].size[1] % blobs[0].size[1] == 0,
outputs.size() == 1, inputs[0].data != outputs[0].data); outputs.size() == 1, inputs[0].data != outputs[0].data);
if (inputs[0].dims == 5) {
CV_Error(Error::StsNotImplemented, "Convolution3D layer is not supported on OCV backend");
}
int ngroups = inputs[0].size[1]/blobs[0].size[1]; int ngroups = inputs[0].size[1]/blobs[0].size[1];
CV_Assert(outputs[0].size[1] % ngroups == 0); CV_Assert(outputs[0].size[1] % ngroups == 0);
int outCn = blobs[0].size[0]; int outCn = blobs[0].size[0];
...@@ -1163,7 +1274,7 @@ public: ...@@ -1163,7 +1274,7 @@ public:
int nstripes = std::max(getNumThreads(), 1); int nstripes = std::max(getNumThreads(), 1);
ParallelConv::run(inputs[0], outputs[0], weightsMat, biasvec, reluslope, ParallelConv::run(inputs[0], outputs[0], weightsMat, biasvec, reluslope,
kernel, pad, stride, dilation, activ.get(), ngroups, nstripes); kernel_size, strides, pads_begin, pads_end, dilations, activ.get(), ngroups, nstripes);
} }
virtual int64 getFLOPS(const std::vector<MatShape> &inputs, virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
...@@ -1172,9 +1283,10 @@ public: ...@@ -1172,9 +1283,10 @@ public:
CV_Assert(inputs.size() == outputs.size()); CV_Assert(inputs.size() == outputs.size());
int64 flops = 0; int64 flops = 0;
int karea = std::accumulate(kernel_size.begin(), kernel_size.end(), 1, std::multiplies<size_t>());
for (int i = 0; i < inputs.size(); i++) for (int i = 0; i < inputs.size(); i++)
{ {
flops += total(outputs[i])*(CV_BIG_INT(2)*kernel.area()*inputs[i][1] + 1); flops += total(outputs[i])*(CV_BIG_INT(2)*karea*inputs[i][1] + 1);
} }
return flops; return flops;
...@@ -1205,29 +1317,39 @@ public: ...@@ -1205,29 +1317,39 @@ public:
virtual bool supportBackend(int backendId) CV_OVERRIDE virtual bool supportBackend(int backendId) CV_OVERRIDE
{ {
#ifdef HAVE_INF_ENGINE #ifdef HAVE_INF_ENGINE
const int outGroupCn = blobs[0].size[1]; // Weights are in IOHW layout const int outGroupCn = blobs[0].size[1]; // Weights are in IOHW or IODHW layout
const int group = numOutput / outGroupCn; const int group = numOutput / outGroupCn;
if (backendId == DNN_BACKEND_INFERENCE_ENGINE) if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
{ {
if (kernel_size.size() == 3) if (kernel_size.size() == 3 && preferableTarget != DNN_TARGET_CPU) {
CV_Error(Error::StsNotImplemented, "Unsupported deconvolution3D layer"); return false;
}
if (adjustPad.height || adjustPad.width) if (std::accumulate(adjust_pads.begin(), adjust_pads.end(), 0, std::plus<size_t>()) > 0)
{ {
if (padMode.empty()) if (padMode.empty())
{ {
if (preferableTarget != DNN_TARGET_CPU && group != 1) if (preferableTarget != DNN_TARGET_CPU && group != 1)
{ {
if ((adjustPad.height && pad.height) || (adjustPad.width && pad.width)) for (int i = 0; i < adjust_pads.size(); i++) {
if (adjust_pads[i] && pads_begin[i])
return false;
}
}
for (int i = 0; i < adjust_pads.size(); i++) {
if (pads_end[i] < adjust_pads[i])
return false; return false;
} }
return pad.width >= adjustPad.width && pad.height >= adjustPad.height; return true;
} }
else if (padMode == "SAME") else if (padMode == "SAME")
{ {
return kernel.width >= pad.width + 1 + adjustPad.width && for (int i = 0; i < adjust_pads.size(); i++) {
kernel.height >= pad.height + 1 + adjustPad.height; if (kernel_size[i] < pads_begin[i] + 1 + adjust_pads[i])
return false;
}
return true;
} }
else if (padMode == "VALID") else if (padMode == "VALID")
return false; return false;
...@@ -1238,7 +1360,7 @@ public: ...@@ -1238,7 +1360,7 @@ public:
return preferableTarget == DNN_TARGET_CPU; return preferableTarget == DNN_TARGET_CPU;
} }
if (preferableTarget == DNN_TARGET_OPENCL || preferableTarget == DNN_TARGET_OPENCL_FP16) if (preferableTarget == DNN_TARGET_OPENCL || preferableTarget == DNN_TARGET_OPENCL_FP16)
return dilation.width == 1 && dilation.height == 1; return std::accumulate(dilations.begin(), dilations.end(), 1, std::multiplies<size_t>()) == 1;
return true; return true;
} }
else else
...@@ -1825,11 +1947,14 @@ public: ...@@ -1825,11 +1947,14 @@ public:
#ifdef HAVE_INF_ENGINE #ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &) CV_OVERRIDE virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &) CV_OVERRIDE
{ {
auto ieWeights = wrapToInfEngineBlob(blobs[0], InferenceEngine::Layout::OIHW); InferenceEngine::Layout layout = blobs[0].dims == 5? InferenceEngine::Layout::NCDHW :
InferenceEngine::Layout::OIHW;
auto ieWeights = wrapToInfEngineBlob(blobs[0], layout);
if (fusedWeights) if (fusedWeights)
{ {
ieWeights = InferenceEngine::make_shared_blob<float>( ieWeights = InferenceEngine::make_shared_blob<float>(
InferenceEngine::Precision::FP32, InferenceEngine::Layout::OIHW, InferenceEngine::Precision::FP32, layout,
ieWeights->dims()); ieWeights->dims());
ieWeights->allocate(); ieWeights->allocate();
...@@ -1838,7 +1963,7 @@ public: ...@@ -1838,7 +1963,7 @@ public:
transpose(weightsMat, newWeights); transpose(weightsMat, newWeights);
} }
const int outGroupCn = blobs[0].size[1]; // Weights are in IOHW layout const int outGroupCn = blobs[0].size[1]; // Weights are in IOHW or OIDHW layout
const int group = numOutput / outGroupCn; const int group = numOutput / outGroupCn;
InferenceEngine::Builder::DeconvolutionLayer ieLayer(name); InferenceEngine::Builder::DeconvolutionLayer ieLayer(name);
...@@ -1850,12 +1975,19 @@ public: ...@@ -1850,12 +1975,19 @@ public:
if (padMode.empty()) if (padMode.empty())
{ {
ieLayer.setPaddingsEnd({pads_end[0] - adjust_pads[0], pads_end[1] - adjust_pads[1]}); std::vector<size_t> paddings_end;
for (int i = 0; i < pads_end.size(); i++) {
paddings_end.push_back(pads_end[i] - adjust_pads[i]);
}
ieLayer.setPaddingsEnd(paddings_end);
} }
else if (padMode == "SAME") else if (padMode == "SAME")
{ {
ieLayer.setPaddingsEnd({kernel_size[0] - pads_begin[0] - 1 - adjust_pads[0], std::vector<size_t> paddings_end;
kernel_size[1] - pads_begin[1] - 1 - adjust_pads[1]}); for (int i = 0; i < pads_begin.size(); i++) {
paddings_end.push_back(kernel_size[i] - pads_begin[i] - 1 - adjust_pads[i]);
}
ieLayer.setPaddingsEnd(paddings_end);
} }
ieLayer.setGroup((size_t)group); ieLayer.setGroup((size_t)group);
ieLayer.setOutDepth((size_t)numOutput); ieLayer.setOutDepth((size_t)numOutput);
...@@ -1875,10 +2007,12 @@ public: ...@@ -1875,10 +2007,12 @@ public:
float flops = 0; float flops = 0;
int outChannels = blobs[0].size[0]; int outChannels = blobs[0].size[0];
size_t karea = std::accumulate(kernel_size.begin(), kernel_size.end(),
1, std::multiplies<size_t>());
for (int i = 0; i < inputs.size(); i++) for (int i = 0; i < inputs.size(); i++)
{ {
flops += CV_BIG_INT(2)*outChannels*kernel.area()*total(inputs[i]); flops += CV_BIG_INT(2)*outChannels*karea*total(inputs[i]);
} }
return flops; return flops;
......
modules/dnn/src/layers/layers_common.cpp
View file @ f6c57388
...@@ -175,11 +175,13 @@ void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kern ...@@ -175,11 +175,13 @@ void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kern
} }
void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin, void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin,
std::vector<size_t>& pads_end, std::vector<size_t>& strides, std::vector<size_t>& dilations, cv::String &padMode) std::vector<size_t>& pads_end, std::vector<size_t>& strides,
std::vector<size_t>& dilations, cv::String &padMode, std::vector<size_t>& adjust_pads)
{ {
util::getKernelSize(params, kernel); util::getKernelSize(params, kernel);
util::getStrideAndPadding(params, pads_begin, pads_end, strides, padMode, kernel.size()); util::getStrideAndPadding(params, pads_begin, pads_end, strides, padMode, kernel.size());
util::getParameter(params, "dilation", "dilation", dilations, true, std::vector<size_t>(kernel.size(), 1)); util::getParameter(params, "dilation", "dilation", dilations, true, std::vector<size_t>(kernel.size(), 1));
util::getParameter(params, "adj", "adj", adjust_pads, true, std::vector<size_t>(kernel.size(), 0));
for (int i = 0; i < dilations.size(); i++) for (int i = 0; i < dilations.size(); i++)
CV_Assert(dilations[i] > 0); CV_Assert(dilations[i] > 0);
......
modules/dnn/src/layers/layers_common.hpp
View file @ f6c57388
...@@ -60,7 +60,8 @@ namespace cv ...@@ -60,7 +60,8 @@ namespace cv
namespace dnn namespace dnn
{ {
void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin, void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin,
std::vector<size_t>& pads_end, std::vector<size_t>& strides, std::vector<size_t>& dilations, cv::String &padMode); std::vector<size_t>& pads_end, std::vector<size_t>& strides, std::vector<size_t>& dilations,
cv::String &padMode, std::vector<size_t>& adjust_pads);
void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kernel, bool &globalPooling, void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kernel, bool &globalPooling,
std::vector<size_t>& pads_begin, std::vector<size_t>& pads_end, std::vector<size_t>& strides, cv::String &padMode); std::vector<size_t>& pads_begin, std::vector<size_t>& pads_end, std::vector<size_t>& strides, cv::String &padMode);
......
modules/dnn/src/layers/pooling_layer.cpp
View file @ f6c57388
...@@ -48,6 +48,7 @@ ...@@ -48,6 +48,7 @@
#include "../op_vkcom.hpp" #include "../op_vkcom.hpp"
#include <float.h> #include <float.h>
#include <algorithm> #include <algorithm>
#include <numeric>
using std::max; using std::max;
using std::min; using std::min;
...@@ -179,13 +180,16 @@ public: ...@@ -179,13 +180,16 @@ public:
} }
else else
{ {
if (!kernel_size.empty() && kernel_size.size() != 2) // TODO Support Pooling3D if (kernel_size.size() == 3)
return (backendId == DNN_BACKEND_OPENCV && preferableTarget == DNN_TARGET_CPU);
if (kernel_size.empty() || kernel_size.size() == 2)
return backendId == DNN_BACKEND_OPENCV ||
(backendId == DNN_BACKEND_HALIDE && haveHalide() &&
(type == MAX || (type == AVE && !pad_t && !pad_l && !pad_b && !pad_r))) ||
(backendId == DNN_BACKEND_VKCOM && haveVulkan() &&
(type == MAX || type == AVE));
else
return false; return false;
return backendId == DNN_BACKEND_OPENCV ||
(backendId == DNN_BACKEND_HALIDE && haveHalide() &&
(type == MAX || (type == AVE && !pad_t && !pad_l && !pad_b && !pad_r))) ||
(backendId == DNN_BACKEND_VKCOM && haveVulkan() &&
(type == MAX || type == AVE));
} }
} }
...@@ -383,19 +387,26 @@ public: ...@@ -383,19 +387,26 @@ public:
int poolingType; int poolingType;
float spatialScale; float spatialScale;
std::vector<size_t> pads_begin, pads_end;
std::vector<size_t> kernel_size;
std::vector<size_t> strides;
PoolingInvoker() : src(0), rois(0), dst(0), mask(0), pad_l(0), pad_t(0), pad_r(0), pad_b(0), PoolingInvoker() : src(0), rois(0), dst(0), mask(0), pad_l(0), pad_t(0), pad_r(0), pad_b(0),
avePoolPaddedArea(false), nstripes(0), avePoolPaddedArea(false), nstripes(0),
computeMaxIdx(0), poolingType(MAX), spatialScale(0) {} computeMaxIdx(0), poolingType(MAX), spatialScale(0) {}
static void run(const Mat& src, const Mat& rois, Mat& dst, Mat& mask, Size kernel, static void run(const Mat& src, const Mat& rois, Mat& dst, Mat& mask,
Size stride, int pad_l, int pad_t, int pad_r, int pad_b, bool avePoolPaddedArea, int poolingType, float spatialScale, std::vector<size_t> kernel_size, std::vector<size_t> strides,
std::vector<size_t> pads_begin, std::vector<size_t> pads_end,
bool avePoolPaddedArea, int poolingType, float spatialScale,
bool computeMaxIdx, int nstripes) bool computeMaxIdx, int nstripes)
{ {
CV_Assert_N( CV_Assert_N(
src.isContinuous(), dst.isContinuous(), src.isContinuous(), dst.isContinuous(),
src.type() == CV_32F, src.type() == dst.type(), src.type() == CV_32F, src.type() == dst.type(),
src.dims == 4, dst.dims == 4, src.dims == 4 || src.dims == 5, dst.dims == 4 || dst.dims == 5,
(((poolingType == ROI || poolingType == PSROI) && dst.size[0] == rois.size[0]) || src.size[0] == dst.size[0]), (((poolingType == ROI || poolingType == PSROI) &&
dst.size[0] == rois.size[0]) || src.size[0] == dst.size[0]),
poolingType == PSROI || src.size[1] == dst.size[1], poolingType == PSROI || src.size[1] == dst.size[1],
(mask.empty() || (mask.type() == src.type() && mask.size == dst.size))); (mask.empty() || (mask.type() == src.type() && mask.size == dst.size)));
...@@ -404,13 +415,20 @@ public: ...@@ -404,13 +415,20 @@ public:
p.src = &src; p.src = &src;
p.rois = &rois; p.rois = &rois;
p.dst = &dst; p.dst = &dst;
p.kernel_size = kernel_size;
p.strides = strides;
p.pads_begin = pads_begin;
p.pads_end = pads_end;
p.mask = &mask; p.mask = &mask;
p.kernel = kernel; p.kernel = Size(kernel_size[1], kernel_size[0]);
p.stride = stride; p.stride = Size(strides[1], strides[0]);
p.pad_l = pad_l; p.pad_l = pads_begin.back();
p.pad_t = pad_t; p.pad_t = pads_begin[pads_begin.size() - 2];
p.pad_r = pad_r; p.pad_r = pads_end.back();
p.pad_b = pad_b; p.pad_b = pads_end[pads_end.size() - 2];
p.avePoolPaddedArea = avePoolPaddedArea; p.avePoolPaddedArea = avePoolPaddedArea;
p.nstripes = nstripes; p.nstripes = nstripes;
p.computeMaxIdx = computeMaxIdx; p.computeMaxIdx = computeMaxIdx;
...@@ -419,10 +437,21 @@ public: ...@@ -419,10 +437,21 @@ public:
if( !computeMaxIdx ) if( !computeMaxIdx )
{ {
p.ofsbuf.resize(kernel.width*kernel.height); int height = src.size[src.dims - 2];
for( int i = 0; i < kernel.height; i++ ) int width = src.size[src.dims - 1];
for( int j = 0; j < kernel.width; j++ )
p.ofsbuf[i*kernel.width + j] = src.size[3]*i + j; int kernel_d = (kernel_size.size() == 3) ? kernel_size[0] : 1;
int kernel_h = kernel_size[kernel_size.size() - 2];
int kernel_w = kernel_size.back();
p.ofsbuf.resize(kernel_d * kernel_h * kernel_w);
for (int i = 0; i < kernel_d; ++i) {
for (int j = 0; j < kernel_h; ++j) {
for (int k = 0; k < kernel_w; ++k) {
p.ofsbuf[i * kernel_h * kernel_w + j * kernel_w + k] = width * height * i + width * j + k;
}
}
}
} }
parallel_for_(Range(0, nstripes), p, nstripes); parallel_for_(Range(0, nstripes), p, nstripes);
...@@ -430,14 +459,29 @@ public: ...@@ -430,14 +459,29 @@ public:
void operator()(const Range& r) const CV_OVERRIDE void operator()(const Range& r) const CV_OVERRIDE
{ {
int channels = dst->size[1], width = dst->size[3], height = dst->size[2]; int channels = dst->size[1];
int inp_width = src->size[3], inp_height = src->size[2];
bool isPool2D = src->dims == 4;
int depth = !isPool2D? dst->size[2] : 1;
int height = dst->size[dst->dims - 2];
int width = dst->size[dst->dims - 1];
int inp_depth = !isPool2D? src->size[2] : 1;
int inp_height = src->size[src->dims - 2];
int inp_width = src->size[src->dims - 1];
size_t total = dst->total(); size_t total = dst->total();
size_t stripeSize = (total + nstripes - 1)/nstripes; size_t stripeSize = (total + nstripes - 1)/nstripes;
size_t stripeStart = r.start*stripeSize; size_t stripeStart = r.start*stripeSize;
size_t stripeEnd = std::min(r.end*stripeSize, total); size_t stripeEnd = std::min(r.end*stripeSize, total);
int kernel_w = kernel.width, kernel_h = kernel.height;
int stride_w = stride.width, stride_h = stride.height; int kernel_d = !isPool2D? kernel_size[0] : 1;
int kernel_h = kernel_size[kernel_size.size() - 2];
int kernel_w = kernel_size.back();
int stride_d = !isPool2D? strides[0] : 0;
int stride_h = strides[strides.size() - 2];
int stride_w = strides.back();
bool compMaxIdx = computeMaxIdx; bool compMaxIdx = computeMaxIdx;
#if CV_SIMD128 #if CV_SIMD128
...@@ -456,9 +500,14 @@ public: ...@@ -456,9 +500,14 @@ public:
ofs /= width; ofs /= width;
int y0 = (int)(ofs % height); int y0 = (int)(ofs % height);
ofs /= height; ofs /= height;
int d0 = (int)(ofs % depth);
ofs /= depth;
int c = (int)(ofs % channels); int c = (int)(ofs % channels);
int n = (int)(ofs / channels); int n = (int)(ofs / channels);
int ystart, yend; int ystart, yend;
int dstart = 0, dend = 1;
const float *srcData = 0; const float *srcData = 0;
if (poolingType == ROI) if (poolingType == ROI)
...@@ -488,15 +537,22 @@ public: ...@@ -488,15 +537,22 @@ public:
} }
else else
{ {
int pad_d_begin = (pads_begin.size() == 3) ? pads_begin[0] : 0;
dstart = d0 * stride_d - pad_d_begin;
dend = min(dstart + kernel_d, (int)(inp_depth + pads_end[0]));
ystart = y0 * stride_h - pad_t; ystart = y0 * stride_h - pad_t;
yend = min(ystart + kernel_h, inp_height + pad_b); yend = min(ystart + kernel_h, inp_height + pad_b);
srcData = src->ptr<float>(n, c); srcData = src->ptr<float>(n, c);
} }
int ddelta = dend - dstart;
dstart = max(dstart, 0);
dend = min(dend, inp_depth);
int ydelta = yend - ystart; int ydelta = yend - ystart;
ystart = max(ystart, 0); ystart = max(ystart, 0);
yend = min(yend, inp_height); yend = min(yend, inp_height);
float *dstData = dst->ptr<float>(n, c, y0); float *dstData = &dst->ptr<float>(n, c, d0)[y0 * width];
float *dstMaskData = mask->data ? mask->ptr<float>(n, c, y0) : 0; float *dstMaskData = mask->data ? &mask->ptr<float>(n, c, d0)[y0 * width] : 0;
int delta = std::min((int)(stripeEnd - ofs0), width - x0); int delta = std::min((int)(stripeEnd - ofs0), width - x0);
ofs0 += delta; ofs0 += delta;
...@@ -516,7 +572,7 @@ public: ...@@ -516,7 +572,7 @@ public:
continue; continue;
} }
#if CV_SIMD128 #if CV_SIMD128
if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_l + kernel_w < inp_width ) if( isPool2D && xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_l + kernel_w < inp_width )
{ {
if( compMaxIdx ) if( compMaxIdx )
{ {
...@@ -621,49 +677,51 @@ public: ...@@ -621,49 +677,51 @@ public:
if( compMaxIdx ) if( compMaxIdx )
{ {
int max_index = -1; int max_index = -1;
for (int y = ystart; y < yend; ++y) for (int d = dstart; d < dend; ++d)
for (int x = xstart; x < xend; ++x) for (int y = ystart; y < yend; ++y)
{ for (int x = xstart; x < xend; ++x)
const int index = y * inp_width + x;
float val = srcData[index];
if (val > max_val)
{ {
max_val = val; const int index = d * inp_width * inp_height + y * inp_width + x;
max_index = index; float val = srcData[index];
if (val > max_val)
{
max_val = val;
max_index = index;
}
} }
}
dstData[x0] = max_val; dstData[x0] = max_val;
if (dstMaskData) if (dstMaskData)
dstMaskData[x0] = max_index; dstMaskData[x0] = max_index;
} }
else else
{ {
for (int y = ystart; y < yend; ++y) for (int d = dstart; d < dend; ++d) {
for (int x = xstart; x < xend; ++x) for (int y = ystart; y < yend; ++y) {
{ for (int x = xstart; x < xend; ++x) {
const int index = y * inp_width + x; const int index = d * inp_width * inp_height + y * inp_width + x;
float val = srcData[index]; float val = srcData[index];
max_val = std::max(max_val, val); max_val = std::max(max_val, val);
}
} }
}
dstData[x0] = max_val; dstData[x0] = max_val;
} }
} }
} }
else if (poolingType == AVE) else if (poolingType == AVE)
{ {
for( ; x0 < x1; x0++ ) for( ; x0 < x1; ++x0)
{ {
int xstart = x0 * stride_w - pad_l; int xstart = x0 * stride_w - pad_l;
int xend = min(xstart + kernel_w, inp_width + pad_r); int xend = min(xstart + kernel_w, inp_width + pad_r);
int xdelta = xend - xstart; int xdelta = xend - xstart;
xstart = max(xstart, 0); xstart = max(xstart, 0);
xend = min(xend, inp_width); xend = min(xend, inp_width);
float inv_kernel_area = avePoolPaddedArea ? xdelta * ydelta : ((yend - ystart) * (xend - xstart)); float inv_kernel_area = avePoolPaddedArea ? xdelta * ydelta * ddelta :
((dend - dstart) * (yend - ystart) * (xend - xstart));
inv_kernel_area = 1.0 / inv_kernel_area; inv_kernel_area = 1.0 / inv_kernel_area;
#if CV_SIMD128 #if CV_SIMD128
if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_l + kernel_w < inp_width ) if( isPool2D && xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_l + kernel_w < inp_width )
{ {
v_float32x4 sum_val0 = v_setzero_f32(), sum_val1 = v_setzero_f32(); v_float32x4 sum_val0 = v_setzero_f32(), sum_val1 = v_setzero_f32();
v_float32x4 ikarea = v_setall_f32(inv_kernel_area); v_float32x4 ikarea = v_setall_f32(inv_kernel_area);
...@@ -689,14 +747,15 @@ public: ...@@ -689,14 +747,15 @@ public:
#endif #endif
{ {
float sum_val = 0.f; float sum_val = 0.f;
for (int y = ystart; y < yend; ++y) for (int d = dstart; d < dend; ++d) {
for (int x = xstart; x < xend; ++x) for (int y = ystart; y < yend; ++y) {
{ for (int x = xstart; x < xend; ++x) {
const int index = y * inp_width + x; const int index = d * inp_width * inp_height + y * inp_width + x;
float val = srcData[index]; float val = srcData[index];
sum_val += val; sum_val += val;
}
} }
}
dstData[x0] = sum_val*inv_kernel_area; dstData[x0] = sum_val*inv_kernel_area;
} }
} }
...@@ -772,21 +831,25 @@ public: ...@@ -772,21 +831,25 @@ public:
{ {
const int nstripes = getNumThreads(); const int nstripes = getNumThreads();
Mat rois; Mat rois;
PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad_l, pad_t, pad_r, pad_b, avePoolPaddedArea, type, spatialScale, computeMaxIdx, nstripes); PoolingInvoker::run(src, rois, dst, mask, kernel_size, strides, pads_begin, pads_end, avePoolPaddedArea, type, spatialScale, computeMaxIdx, nstripes);
} }
void avePooling(Mat &src, Mat &dst) void avePooling(Mat &src, Mat &dst)
{ {
const int nstripes = getNumThreads(); const int nstripes = getNumThreads();
Mat rois, mask; Mat rois, mask;
PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad_l, pad_t, pad_r, pad_b, avePoolPaddedArea, type, spatialScale, computeMaxIdx, nstripes); PoolingInvoker::run(src, rois, dst, mask, kernel_size, strides, pads_begin, pads_end, avePoolPaddedArea, type, spatialScale, computeMaxIdx, nstripes);
} }
void roiPooling(const Mat &src, const Mat &rois, Mat &dst) void roiPooling(const Mat &src, const Mat &rois, Mat &dst)
{ {
const int nstripes = getNumThreads(); const int nstripes = getNumThreads();
Mat mask; Mat mask;
PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad_l, pad_t, pad_r, pad_b, avePoolPaddedArea, type, spatialScale, computeMaxIdx, nstripes); kernel_size.resize(2);
strides.resize(2);
pads_begin.resize(2);
pads_end.resize(2);
PoolingInvoker::run(src, rois, dst, mask, kernel_size, strides, pads_begin, pads_end, avePoolPaddedArea, type, spatialScale, computeMaxIdx, nstripes);
} }
virtual Ptr<BackendNode> initMaxPoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs) virtual Ptr<BackendNode> initMaxPoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
...@@ -974,17 +1037,18 @@ public: ...@@ -974,17 +1037,18 @@ public:
{ {
CV_UNUSED(inputs); // suppress unused variable warning CV_UNUSED(inputs); // suppress unused variable warning
long flops = 0; long flops = 0;
size_t karea = std::accumulate(kernel_size.begin(), kernel_size.end(),
1, std::multiplies<size_t>());
for(int i = 0; i < outputs.size(); i++) for(int i = 0; i < outputs.size(); i++)
{ {
if (type == MAX) if (type == MAX)
{ {
if (i%2 == 0) if (i%2 == 0)
flops += total(outputs[i])*kernel.area(); flops += total(outputs[i])*karea;
} }
else else
{ {
flops += total(outputs[i])*(kernel.area() + 1); flops += total(outputs[i])*(karea + 1);
} }
} }
return flops; return flops;
......
modules/dnn/src/onnx/onnx_importer.cpp
View file @ f6c57388
...@@ -682,42 +682,37 @@ void ONNXImporter::populateNet(Net dstNet) ...@@ -682,42 +682,37 @@ void ONNXImporter::populateNet(Net dstNet)
layerParams.set("num_output", layerParams.blobs[0].size[1] * layerParams.get<int>("group", 1)); layerParams.set("num_output", layerParams.blobs[0].size[1] * layerParams.get<int>("group", 1));
layerParams.set("bias_term", node_proto.input_size() == 3); layerParams.set("bias_term", node_proto.input_size() == 3);
if (!layerParams.has("kernel_size"))
CV_Error(Error::StsNotImplemented,
"Required attribute 'kernel_size' is not present.");
if (layerParams.has("output_shape")) if (layerParams.has("output_shape"))
{ {
const DictValue& outShape = layerParams.get("output_shape"); const DictValue& outShape = layerParams.get("output_shape");
DictValue strides = layerParams.get("stride");
DictValue kernel = layerParams.get("kernel_size");
if (outShape.size() != 4) String padMode;
CV_Error(Error::StsNotImplemented, "Output shape must have 4 elements."); std::vector<int> adjust_pads;
if (layerParams.has("pad_mode"))
DictValue stride = layerParams.get("stride");
const int strideY = stride.getIntValue(0);
const int strideX = stride.getIntValue(1);
const int outH = outShape.getIntValue(2);
const int outW = outShape.getIntValue(3);
if (layerParams.get<String>("pad_mode") == "SAME")
{ {
layerParams.set("adj_w", (outW - 1) % strideX); padMode = toUpperCase(layerParams.get<String>("pad_mode"));
layerParams.set("adj_h", (outH - 1) % strideY); if (padMode != "SAME" && padMode != "VALID")
} CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
else if (layerParams.get<String>("pad_mode") == "VALID")
{ for (int i = 0; i < strides.size(); i++)
if (!layerParams.has("kernel_size")) {
CV_Error(Error::StsNotImplemented, int sz = outShape.get<int>(2 + i);
"Required attribute 'kernel_size' is not present."); int stride = strides.get<int>(i);
adjust_pads.push_back(padMode == "SAME"? (sz - 1) % stride :
DictValue kernel = layerParams.get("kernel_size"); (sz - kernel.get<int>(i)) % stride);
layerParams.set("adj_h", (outH - kernel.getIntValue(0)) % strideY); }
layerParams.set("adj_w", (outW - kernel.getIntValue(1)) % strideX); layerParams.set("adj", DictValue::arrayInt(&adjust_pads[0], adjust_pads.size()));
} }
} }
else if (layerParams.has("output_padding")) else if (layerParams.has("output_padding"))
{ {
const DictValue& adj_pad = layerParams.get("output_padding"); replaceLayerParam(layerParams, "output_padding", "adj");
if (adj_pad.size() != 2)
CV_Error(Error::StsNotImplemented, "Deconvolution3D layer is not supported");
layerParams.set("adj_w", adj_pad.get<int>(1));
layerParams.set("adj_h", adj_pad.get<int>(0));
} }
} }
else if (layer_type == "Transpose") else if (layer_type == "Transpose")
......
modules/dnn/test/test_onnx_importer.cpp
View file @ f6c57388
...@@ -100,8 +100,8 @@ TEST_P(Test_ONNX_layers, Convolution3D) ...@@ -100,8 +100,8 @@ TEST_P(Test_ONNX_layers, Convolution3D)
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1"); throw SkipTestException("Test is enabled starts from 2019R1");
#endif #endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU) if (target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported"); throw SkipTestException("Only CPU is supported");
testONNXModels("conv3d"); testONNXModels("conv3d");
testONNXModels("conv3d_bias"); testONNXModels("conv3d_bias");
} }
...@@ -127,6 +127,19 @@ TEST_P(Test_ONNX_layers, Deconvolution) ...@@ -127,6 +127,19 @@ TEST_P(Test_ONNX_layers, Deconvolution)
testONNXModels("deconv_adjpad_2d", npy, 0, 0, false, false); testONNXModels("deconv_adjpad_2d", npy, 0, 0, false, false);
} }
TEST_P(Test_ONNX_layers, Deconvolution3D)
{
#if defined(INF_ENGINE_RELEASE)
applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_2018R5);
#endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported");
testONNXModels("deconv3d");
testONNXModels("deconv3d_bias");
testONNXModels("deconv3d_pad");
testONNXModels("deconv3d_adjpad");
}
TEST_P(Test_ONNX_layers, Dropout) TEST_P(Test_ONNX_layers, Dropout)
{ {
testONNXModels("dropout"); testONNXModels("dropout");
...@@ -185,8 +198,8 @@ TEST_P(Test_ONNX_layers, MaxPooling3D) ...@@ -185,8 +198,8 @@ TEST_P(Test_ONNX_layers, MaxPooling3D)
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1"); throw SkipTestException("Test is enabled starts from 2019R1");
#endif #endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU) if (target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported"); throw SkipTestException("Only CPU is supported");
testONNXModels("max_pool3d"); testONNXModels("max_pool3d");
} }
...@@ -195,11 +208,21 @@ TEST_P(Test_ONNX_layers, AvePooling3D) ...@@ -195,11 +208,21 @@ TEST_P(Test_ONNX_layers, AvePooling3D)
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1"); throw SkipTestException("Test is enabled starts from 2019R1");
#endif #endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU) if (target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported"); throw SkipTestException("Only CPU is supported");
testONNXModels("ave_pool3d"); testONNXModels("ave_pool3d");
} }
TEST_P(Test_ONNX_layers, PoolConv3D)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1");
#endif
if (target != DNN_TARGET_CPU)
throw SkipTestException("Only CPU is supported");
testONNXModels("pool_conv_3d");
}
TEST_P(Test_ONNX_layers, BatchNormalization) TEST_P(Test_ONNX_layers, BatchNormalization)
{ {
testONNXModels("batch_norm"); testONNXModels("batch_norm");
...@@ -579,10 +602,10 @@ TEST_P(Test_ONNX_nets, Resnet34_kinetics) ...@@ -579,10 +602,10 @@ TEST_P(Test_ONNX_nets, Resnet34_kinetics)
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1"); throw SkipTestException("Test is enabled starts from 2019R1");
#endif #endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU) if (target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported"); throw SkipTestException("Only CPU is supported");
String onnxmodel = findDataFile("dnn/resnet-34_kinetics.onnx", false); String onnxmodel = findDataFile("dnn/resnet-34_kinetics.onnx");
Mat image0 = imread(findDataFile("dnn/dog416.png")); Mat image0 = imread(findDataFile("dnn/dog416.png"));
Mat image1 = imread(findDataFile("dnn/street.png")); Mat image1 = imread(findDataFile("dnn/street.png"));
......
modules/dnn/test/test_tf_importer.cpp
View file @ f6c57388
...@@ -136,8 +136,8 @@ TEST_P(Test_TensorFlow_layers, Convolution3D) ...@@ -136,8 +136,8 @@ TEST_P(Test_TensorFlow_layers, Convolution3D)
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1"); throw SkipTestException("Test is enabled starts from 2019R1");
#endif #endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU) if (target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported"); throw SkipTestException("Only CPU is supported");
runTensorFlowNet("conv3d"); runTensorFlowNet("conv3d");
} }
...@@ -243,8 +243,8 @@ TEST_P(Test_TensorFlow_layers, MaxPooling3D) ...@@ -243,8 +243,8 @@ TEST_P(Test_TensorFlow_layers, MaxPooling3D)
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1"); throw SkipTestException("Test is enabled starts from 2019R1");
#endif #endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU) if (target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported"); throw SkipTestException("Only CPU is supported");
runTensorFlowNet("max_pool3d"); runTensorFlowNet("max_pool3d");
} }
...@@ -253,8 +253,8 @@ TEST_P(Test_TensorFlow_layers, AvePooling3D) ...@@ -253,8 +253,8 @@ TEST_P(Test_TensorFlow_layers, AvePooling3D)
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1"); throw SkipTestException("Test is enabled starts from 2019R1");
#endif #endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU) if (target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported"); throw SkipTestException("Only CPU is supported");
runTensorFlowNet("ave_pool3d"); runTensorFlowNet("ave_pool3d");
} }
......
modules/imgproc/src/filter.simd.hpp
View file @ f6c57388
...@@ -84,6 +84,7 @@ Ptr<BaseFilter> getLinearFilter( ...@@ -84,6 +84,7 @@ Ptr<BaseFilter> getLinearFilter(
#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY #ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
typedef int CV_DECL_ALIGNED(1) unaligned_int;
#define VEC_ALIGN CV_MALLOC_ALIGN #define VEC_ALIGN CV_MALLOC_ALIGN
int FilterEngine__start(FilterEngine& this_, const Size &_wholeSize, const Size &sz, const Point &ofs) int FilterEngine__start(FilterEngine& this_, const Size &_wholeSize, const Size &sz, const Point &ofs)
...@@ -1049,7 +1050,7 @@ struct SymmColumnVec_32s8u ...@@ -1049,7 +1050,7 @@ struct SymmColumnVec_32s8u
s0 = v_muladd(v_cvt_f32(v_load(src[k] + i) + v_load(src[-k] + i)), v_setall_f32(ky[k]), s0); s0 = v_muladd(v_cvt_f32(v_load(src[k] + i) + v_load(src[-k] + i)), v_setall_f32(ky[k]), s0);
v_int32x4 s32 = v_round(s0); v_int32x4 s32 = v_round(s0);
v_int16x8 s16 = v_pack(s32, s32); v_int16x8 s16 = v_pack(s32, s32);
*(int*)(dst + i) = v_reinterpret_as_s32(v_pack_u(s16, s16)).get0(); *(unaligned_int*)(dst + i) = v_reinterpret_as_s32(v_pack_u(s16, s16)).get0();
i += v_int32x4::nlanes; i += v_int32x4::nlanes;
} }
} }
...@@ -1104,7 +1105,7 @@ struct SymmColumnVec_32s8u ...@@ -1104,7 +1105,7 @@ struct SymmColumnVec_32s8u
s0 = v_muladd(v_cvt_f32(v_load(src[k] + i) - v_load(src[-k] + i)), v_setall_f32(ky[k]), s0); s0 = v_muladd(v_cvt_f32(v_load(src[k] + i) - v_load(src[-k] + i)), v_setall_f32(ky[k]), s0);
v_int32x4 s32 = v_round(s0); v_int32x4 s32 = v_round(s0);
v_int16x8 s16 = v_pack(s32, s32); v_int16x8 s16 = v_pack(s32, s32);
*(int*)(dst + i) = v_reinterpret_as_s32(v_pack_u(s16, s16)).get0(); *(unaligned_int*)(dst + i) = v_reinterpret_as_s32(v_pack_u(s16, s16)).get0();
i += v_int32x4::nlanes; i += v_int32x4::nlanes;
} }
} }
...@@ -2129,7 +2130,7 @@ struct FilterVec_8u ...@@ -2129,7 +2130,7 @@ struct FilterVec_8u
s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_load_expand_q(src[k] + i))), v_setall_f32(kf[k]), s0); s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_load_expand_q(src[k] + i))), v_setall_f32(kf[k]), s0);
v_int32x4 s32 = v_round(s0); v_int32x4 s32 = v_round(s0);
v_int16x8 s16 = v_pack(s32, s32); v_int16x8 s16 = v_pack(s32, s32);
*(int*)(dst + i) = v_reinterpret_as_s32(v_pack_u(s16, s16)).get0(); *(unaligned_int*)(dst + i) = v_reinterpret_as_s32(v_pack_u(s16, s16)).get0();
i += v_int32x4::nlanes; i += v_int32x4::nlanes;
} }
return i; return i;
......
modules/imgproc/src/smooth.simd.hpp
View file @ f6c57388
...@@ -334,7 +334,7 @@ void hlineSmooth3Naba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ...@@ -334,7 +334,7 @@ void hlineSmooth3Naba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const
{ {
int src_idx = borderInterpolate(-1, len, borderType); int src_idx = borderInterpolate(-1, len, borderType);
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
((uint16_t*)dst)[k] = ((uint16_t*)m)[1] * src[k] + ((uint16_t*)m)[0] * ((uint16_t)(src[cn + k]) + (uint16_t)(src[src_idx*cn + k])); ((uint16_t*)dst)[k] = saturate_cast<uint16_t>(((uint16_t*)m)[1] * (uint32_t)(src[k]) + ((uint16_t*)m)[0] * ((uint32_t)(src[cn + k]) + (uint32_t)(src[src_idx*cn + k])));
} }
else else
{ {
...@@ -354,14 +354,14 @@ void hlineSmooth3Naba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ...@@ -354,14 +354,14 @@ void hlineSmooth3Naba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const
v_mul_wrap(vx_load_expand(src), v_mul1)); v_mul_wrap(vx_load_expand(src), v_mul1));
#endif #endif
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
*((uint16_t*)dst) = ((uint16_t*)m)[1] * src[0] + ((uint16_t*)m)[0] * ((uint16_t)(src[-cn]) + (uint16_t)(src[cn])); *((uint16_t*)dst) = saturate_cast<uint16_t>(((uint16_t*)m)[1] * (uint32_t)(src[0]) + ((uint16_t*)m)[0] * ((uint32_t)(src[-cn]) + (uint32_t)(src[cn])));
// Point that fall right from border // Point that fall right from border
if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped
{ {
int src_idx = (borderInterpolate(len, len, borderType) - (len - 1))*cn; int src_idx = (borderInterpolate(len, len, borderType) - (len - 1))*cn;
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
((uint16_t*)dst)[k] = ((uint16_t*)m)[1] * src[k] + ((uint16_t*)m)[0] * ((uint16_t)(src[k - cn]) + (uint16_t)(src[src_idx + k])); ((uint16_t*)dst)[k] = saturate_cast<uint16_t>(((uint16_t*)m)[1] * (uint32_t)(src[k]) + ((uint16_t*)m)[0] * ((uint32_t)(src[k - cn]) + (uint32_t)(src[src_idx + k])));
} }
else else
{ {
...@@ -896,8 +896,8 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -896,8 +896,8 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
int idxp2 = borderInterpolate(3, len, borderType)*cn; int idxp2 = borderInterpolate(3, len, borderType)*cn;
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
{ {
((uint16_t*)dst)[k] = ((uint16_t*)m)[1] * ((uint16_t)(src[k + idxm1]) + (uint16_t)(src[k + cn])) + ((uint16_t*)m)[2] * src[k] + ((uint16_t*)m)[0] * ((uint16_t)(src[k + idxp1]) + (uint16_t)(src[k + idxm2])); ((uint16_t*)dst)[k] = saturate_cast<uint16_t>(((uint16_t*)m)[1] * ((uint32_t)(src[k + idxm1]) + (uint32_t)(src[k + cn])) + ((uint16_t*)m)[2] * (uint32_t)(src[k]) + ((uint16_t*)m)[0] * ((uint32_t)(src[k + idxp1]) + (uint32_t)(src[k + idxm2])));
((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[0] * ((uint16_t)(src[k + idxm1]) + (uint16_t)(src[k + idxp2])) + ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[k + idxp1])) + ((uint16_t*)m)[2] * src[k + cn]; ((uint16_t*)dst)[k + cn] = saturate_cast<uint16_t>(((uint16_t*)m)[0] * ((uint32_t)(src[k + idxm1]) + (uint32_t)(src[k + idxp2])) + ((uint16_t*)m)[1] * ((uint32_t)(src[k]) + (uint32_t)(src[k + idxp1])) + ((uint16_t*)m)[2] * (uint32_t)(src[k + cn]));
} }
} }
} }
...@@ -907,7 +907,7 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -907,7 +907,7 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
{ {
dst[k] = m[2] * src[k] + m[1] * src[k + cn] + m[0] * src[k + 2 * cn]; dst[k] = m[2] * src[k] + m[1] * src[k + cn] + m[0] * src[k + 2 * cn];
((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[k + 2 * cn])) + ((uint16_t*)m)[2] * src[k + cn]; ((uint16_t*)dst)[k + cn] = saturate_cast<uint16_t>(((uint16_t*)m)[1] * ((uint32_t)(src[k]) + (uint32_t)(src[k + 2 * cn])) + ((uint16_t*)m)[2] * (uint32_t)(src[k + cn]));
dst[k + 2 * cn] = m[0] * src[k] + m[1] * src[k + cn] + m[2] * src[k + 2 * cn]; dst[k + 2 * cn] = m[0] * src[k] + m[1] * src[k + cn] + m[2] * src[k + 2 * cn];
} }
else else
...@@ -918,9 +918,9 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -918,9 +918,9 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
int idxp2 = borderInterpolate(4, len, borderType)*cn; int idxp2 = borderInterpolate(4, len, borderType)*cn;
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
{ {
((uint16_t*)dst)[k] = ((uint16_t*)m)[2] * src[k] + ((uint16_t*)m)[1] * ((uint16_t)(src[k + cn]) + (uint16_t)(src[k + idxm1])) + ((uint16_t*)m)[0] * ((uint16_t)(src[k + 2 * cn]) + (uint16_t)(src[k + idxm2])); ((uint16_t*)dst)[k] = saturate_cast<uint16_t>(((uint16_t*)m)[2] * (uint32_t)(src[k]) + ((uint16_t*)m)[1] * ((uint32_t)(src[k + cn]) + (uint32_t)(src[k + idxm1])) + ((uint16_t*)m)[0] * ((uint32_t)(src[k + 2 * cn]) + (uint32_t)(src[k + idxm2])));
((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[2] * src[k + cn] + ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[k + 2 * cn])) + ((uint16_t*)m)[0] * ((uint16_t)(src[k + idxm1]) + (uint16_t)(src[k + idxp1])); ((uint16_t*)dst)[k + cn] = saturate_cast<uint16_t>(((uint16_t*)m)[2] * (uint32_t)(src[k + cn]) + ((uint16_t*)m)[1] * ((uint32_t)(src[k]) + (uint32_t)(src[k + 2 * cn])) + ((uint16_t*)m)[0] * ((uint32_t)(src[k + idxm1]) + (uint32_t)(src[k + idxp1])));
((uint16_t*)dst)[k + 2 * cn] = ((uint16_t*)m)[0] * ((uint16_t)(src[k]) + (uint16_t)(src[k + idxp2])) + ((uint16_t*)m)[1] * ((uint16_t)(src[k + cn]) + (uint16_t)(src[k + idxp1])) + ((uint16_t*)m)[2] * src[k + 2 * cn]; ((uint16_t*)dst)[k + 2 * cn] = saturate_cast<uint16_t>(((uint16_t*)m)[0] * ((uint32_t)(src[k]) + (uint32_t)(src[k + idxp2])) + ((uint16_t*)m)[1] * ((uint32_t)(src[k + cn]) + (uint32_t)(src[k + idxp1])) + ((uint16_t*)m)[2] * (uint32_t)(src[k + 2 * cn]));
} }
} }
} }
...@@ -933,8 +933,8 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -933,8 +933,8 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
int idxm1 = borderInterpolate(-1, len, borderType)*cn; int idxm1 = borderInterpolate(-1, len, borderType)*cn;
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
{ {
((uint16_t*)dst)[k] = ((uint16_t*)m)[2] * src[k] + ((uint16_t*)m)[1] * ((uint16_t)(src[cn + k]) + (uint16_t)(src[idxm1 + k])) + ((uint16_t*)m)[0] * ((uint16_t)(src[2 * cn + k]) + (uint16_t)(src[idxm2 + k])); ((uint16_t*)dst)[k] = saturate_cast<uint16_t>(((uint16_t*)m)[2] * (uint32_t)(src[k]) + ((uint16_t*)m)[1] * ((uint32_t)(src[cn + k]) + (uint32_t)(src[idxm1 + k])) + ((uint16_t*)m)[0] * ((uint32_t)(src[2 * cn + k]) + (uint32_t)(src[idxm2 + k])));
((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[2 * cn + k])) + ((uint16_t*)m)[2] * src[cn + k] + ((uint16_t*)m)[0] * ((uint16_t)(src[3 * cn + k]) + (uint16_t)(src[idxm1 + k])); ((uint16_t*)dst)[k + cn] = saturate_cast<uint16_t>(((uint16_t*)m)[1] * ((uint32_t)(src[k]) + (uint32_t)(src[2 * cn + k])) + ((uint16_t*)m)[2] * (uint32_t)(src[cn + k]) + ((uint16_t*)m)[0] * ((uint32_t)(src[3 * cn + k]) + (uint32_t)(src[idxm1 + k])));
} }
} }
else else
...@@ -942,7 +942,7 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -942,7 +942,7 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
{ {
dst[k] = m[2] * src[k] + m[1] * src[cn + k] + m[0] * src[2 * cn + k]; dst[k] = m[2] * src[k] + m[1] * src[cn + k] + m[0] * src[2 * cn + k];
((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[2 * cn + k])) + ((uint16_t*)m)[2] * src[cn + k] + ((uint16_t*)m)[0] * src[3 * cn + k]; ((uint16_t*)dst)[k + cn] = saturate_cast<uint16_t>(((uint16_t*)m)[1] * ((uint32_t)(src[k]) + (uint32_t)(src[2 * cn + k])) + ((uint16_t*)m)[2] * (uint32_t)(src[cn + k]) + ((uint16_t*)m)[0] * (uint32_t)(src[3 * cn + k]));
} }
} }
...@@ -960,7 +960,7 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -960,7 +960,7 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
v_mul_wrap(vx_load_expand(src), v_mul2)); v_mul_wrap(vx_load_expand(src), v_mul2));
#endif #endif
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
*((uint16_t*)dst) = ((uint16_t*)m)[0] * ((uint16_t)(src[-2 * cn]) + (uint16_t)(src[2 * cn])) + ((uint16_t*)m)[1] * ((uint16_t)(src[-cn]) + (uint16_t)(src[cn])) + ((uint16_t*)m)[2] * src[0]; *((uint16_t*)dst) = saturate_cast<uint16_t>(((uint16_t*)m)[0] * ((uint32_t)(src[-2 * cn]) + (uint32_t)(src[2 * cn])) + ((uint16_t*)m)[1] * ((uint32_t)(src[-cn]) + (uint32_t)(src[cn])) + ((uint16_t*)m)[2] * (uint32_t)(src[0]));
// Points that fall right from border // Points that fall right from border
if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped
...@@ -969,15 +969,15 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -969,15 +969,15 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
int idxp2 = (borderInterpolate(len + 1, len, borderType) - (len - 2))*cn; int idxp2 = (borderInterpolate(len + 1, len, borderType) - (len - 2))*cn;
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
{ {
((uint16_t*)dst)[k] = ((uint16_t*)m)[0] * ((uint16_t)(src[k - 2 * cn]) + (uint16_t)(src[idxp1 + k])) + ((uint16_t*)m)[1] * ((uint16_t)(src[k - cn]) + (uint16_t)(src[k + cn])) + ((uint16_t*)m)[2] * src[k]; ((uint16_t*)dst)[k] = saturate_cast<uint16_t>(((uint16_t*)m)[0] * ((uint32_t)(src[k - 2 * cn]) + (uint32_t)(src[idxp1 + k])) + ((uint16_t*)m)[1] * ((uint32_t)(src[k - cn]) + (uint32_t)(src[k + cn])) + ((uint16_t*)m)[2] * (uint32_t)(src[k]));
((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[0] * ((uint16_t)(src[k - cn]) + (uint16_t)(src[idxp2 + k])) + ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[idxp1 + k])) + ((uint16_t*)m)[2] * src[k + cn]; ((uint16_t*)dst)[k + cn] = saturate_cast<uint16_t>(((uint16_t*)m)[0] * ((uint32_t)(src[k - cn]) + (uint32_t)(src[idxp2 + k])) + ((uint16_t*)m)[1] * ((uint32_t)(src[k]) + (uint32_t)(src[idxp1 + k])) + ((uint16_t*)m)[2] * (uint32_t)(src[k + cn]));
} }
} }
else else
{ {
for (int k = 0; k < cn; k++) for (int k = 0; k < cn; k++)
{ {
((uint16_t*)dst)[k] = ((uint16_t*)m)[0] * src[k - 2 * cn] + ((uint16_t*)m)[1] * ((uint16_t)(src[k - cn]) + (uint16_t)(src[k + cn])) + ((uint16_t*)m)[2] * src[k]; ((uint16_t*)dst)[k] = saturate_cast<uint16_t>(((uint16_t*)m)[0] * (uint32_t)(src[k - 2 * cn]) + ((uint16_t*)m)[1] * ((uint32_t)(src[k - cn]) + (uint32_t)(src[k + cn])) + ((uint16_t*)m)[2] * (uint32_t)(src[k]));
dst[k + cn] = m[0] * src[k - cn] + m[1] * src[k] + m[2] * src[k + cn]; dst[k + cn] = m[0] * src[k - cn] + m[1] * src[k] + m[2] * src[k + cn];
} }
} }
......
modules/imgproc/test/test_smooth_bitexact.cpp
View file @ f6c57388
...@@ -158,4 +158,12 @@ TEST(GaussianBlur_Bitexact, Linear8U) ...@@ -158,4 +158,12 @@ TEST(GaussianBlur_Bitexact, Linear8U)
} }
} }
TEST(GaussianBlur_Bitexact, regression_15015)
{
Mat src(100,100,CV_8UC3,Scalar(255,255,255));
Mat dst;
GaussianBlur(src, dst, Size(5, 5), 9);
ASSERT_EQ(0.0, cvtest::norm(dst, src, NORM_INF));
}
}} // namespace }} // namespace
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment