Merge pull request #21703 from rogday:transpose

Add n-dimensional transpose to core

* add n-dimensional transpose to core

* add performance test, write sequentially and address review comments

modules/core/include/opencv2/core.hpp

@@ -1739,6 +1739,16 @@ should be done separately if needed.
 */
 CV_EXPORTS_W void transpose(InputArray src, OutputArray dst);
 
+/** @brief Transpose for n-dimensional matrices.
+ *
+ * @note Input should be continuous single-channel matrix.
+ * @param src input array.
+ * @param order a permutation of [0,1,..,N-1] where N is the number of axes of src.
+ * The i’th axis of dst will correspond to the axis numbered order[i] of the input.
+ * @param dst output array of the same type as src.
+ */
+CV_EXPORTS_W void transposeND(InputArray src, const std::vector<int>& order, OutputArray dst);
+
 /** @brief Performs the matrix transformation of every array element.
 
 The function cv::transform performs the matrix transformation of every

modules/core/perf/perf_arithm.cpp

 #include "perf_precomp.hpp"
+#include <numeric>
 
 namespace opencv_test
 {
@@ -393,6 +394,29 @@ PERF_TEST_P_(BinaryOpTest, reciprocal)
     SANITY_CHECK_NOTHING();
 }
 
+
+PERF_TEST_P_(BinaryOpTest, transposeND)
+{
+    Size sz = get<0>(GetParam());
+    int type = get<1>(GetParam());
+    cv::Mat a = Mat(sz, type).reshape(1);
+
+    std::vector<int> order(a.dims);
+    std::iota(order.begin(), order.end(), 0);
+    std::reverse(order.begin(), order.end());
+
+    std::vector<int> new_sz(a.dims);
+    std::copy(a.size.p, a.size.p + a.dims, new_sz.begin());
+    std::reverse(new_sz.begin(), new_sz.end());
+    cv::Mat b = Mat(new_sz, type);
+
+    declare.in(a,WARMUP_RNG).out(b);
+
+    TEST_CYCLE() cv::transposeND(a, order, b);
+
+    SANITY_CHECK_NOTHING();
+}
+
 INSTANTIATE_TEST_CASE_P(/*nothing*/ , BinaryOpTest,
     testing::Combine(
         testing::Values(szVGA, sz720p, sz1080p),

modules/core/src/matrix_transform.cpp

@@ -4,6 +4,7 @@
 
 #include "precomp.hpp"
 #include "opencl_kernels_core.hpp"
+#include "opencv2/core/detail/dispatch_helper.impl.hpp"
 
 namespace cv {
 
@@ -282,6 +283,72 @@ void transpose( InputArray _src, OutputArray _dst )
 }
 
 
+void transposeND(InputArray src_, const std::vector<int>& order, OutputArray dst_)
+{
+    Mat inp = src_.getMat();
+    CV_Assert(inp.isContinuous());
+    CV_CheckEQ(inp.channels(), 1, "Input array should be single-channel");
+    CV_CheckEQ(order.size(), static_cast<size_t>(inp.dims), "Number of dimensions shouldn't change");
+
+    auto order_ = order;
+    std::sort(order_.begin(), order_.end());
+    for (size_t i = 0; i < order_.size(); ++i)
+    {
+        CV_CheckEQ(static_cast<size_t>(order_[i]), i, "New order should be a valid permutation of the old one");
+    }
+
+    std::vector<int> newShape(order.size());
+    for (size_t i = 0; i < order.size(); ++i)
+    {
+        newShape[i] = inp.size[order[i]];
+    }
+
+    dst_.create(static_cast<int>(newShape.size()), newShape.data(), inp.type());
+    Mat out = dst_.getMat();
+    CV_Assert(out.isContinuous());
+    CV_Assert(inp.data != out.data);
+
+    int continuous_idx = 0;
+    for (int i = static_cast<int>(order.size()) - 1; i >= 0; --i)
+    {
+        if (order[i] != i)
+        {
+            continuous_idx = i + 1;
+            break;
+        }
+    }
+
+    size_t continuous_size = continuous_idx == 0 ? out.total() : out.step1(continuous_idx - 1);
+    size_t outer_size = out.total() / continuous_size;
+
+    std::vector<size_t> steps(order.size());
+    for (int i = 0; i < static_cast<int>(steps.size()); ++i)
+    {
+        steps[i] = inp.step1(order[i]);
+    }
+
+    auto* src = inp.ptr<const unsigned char>();
+    auto* dst = out.ptr<unsigned char>();
+
+    size_t src_offset = 0;
+    size_t es = out.elemSize();
+    for (size_t i = 0; i < outer_size; ++i)
+    {
+        std::memcpy(dst, src + es * src_offset, es * continuous_size);
+        dst += es * continuous_size;
+        for (int j = continuous_idx - 1; j >= 0; --j)
+        {
+            src_offset += steps[j];
+            if ((src_offset / steps[j]) % out.size[j] != 0)
+            {
+                break;
+            }
+            src_offset -= steps[j] * out.size[j];
+        }
+    }
+}
+
+
 #if CV_SIMD128
 template<typename V> CV_ALWAYS_INLINE void flipHoriz_single( const uchar* src, size_t sstep, uchar* dst, size_t dstep, Size size, size_t esz )
 {

modules/core/test/test_arithm.cpp

@@ -3,6 +3,7 @@
 // of this distribution and at http://opencv.org/license.html.
 #include "test_precomp.hpp"
 #include "ref_reduce_arg.impl.hpp"
+#include <algorithm>
 
 namespace opencv_test { namespace {
 
@@ -2128,6 +2129,79 @@ TEST(Core_minMaxIdx, regression_9207_1)
 }
 
 
+class TransposeND : public testing::TestWithParam< tuple<std::vector<int>, perf::MatType> >
+{
+public:
+    std::vector<int> m_shape;
+    int m_type;
+
+    void SetUp()
+    {
+        std::tie(m_shape, m_type) = GetParam();
+    }
+};
+
+
+TEST_P(TransposeND, basic)
+{
+    Mat inp(m_shape, m_type);
+    randu(inp, 0, 255);
+
+    std::vector<int> order(m_shape.size());
+    std::iota(order.begin(), order.end(), 0);
+    auto transposer = [&order] (const std::vector<int>& id)
+    {
+        std::vector<int> ret(id.size());
+        for (size_t i = 0; i < id.size(); ++i)
+        {
+            ret[i] = id[order[i]];
+        }
+        return ret;
+    };
+    auto advancer = [&inp] (std::vector<int>& id)
+    {
+        for (int j = static_cast<int>(id.size() - 1); j >= 0; --j)
+        {
+            ++id[j];
+            if (id[j] != inp.size[j])
+            {
+                break;
+            }
+            id[j] = 0;
+        }
+    };
+
+    do
+    {
+        Mat out;
+        cv::transposeND(inp, order, out);
+        std::vector<int> id(order.size());
+        for (size_t i = 0; i < inp.total(); ++i)
+        {
+            auto new_id = transposer(id);
+            switch (inp.type())
+            {
+            case CV_8UC1:
+                ASSERT_EQ(inp.at<uint8_t>(id.data()), out.at<uint8_t>(new_id.data()));
+                break;
+            case CV_32FC1:
+                ASSERT_EQ(inp.at<float>(id.data()), out.at<float>(new_id.data()));
+                break;
+            default:
+                FAIL() << "Unsupported type: " << inp.type();
+            }
+            advancer(id);
+        }
+    } while (std::next_permutation(order.begin(), order.end()));
+}
+
+
+INSTANTIATE_TEST_CASE_P(Arithm, TransposeND, testing::Combine(
+    testing::Values(std::vector<int>{2, 3, 4}, std::vector<int>{5, 10}),
+    testing::Values(perf::MatType(CV_8UC1), CV_32FC1)
+));
+
+
 TEST(Core_minMaxIdx, regression_9207_2)
 {
     const int rows = 13;