core cuda changes

modules/core/include/opencv2/core/cuda.hpp

@@ -158,6 +158,10 @@ public:
     //! decreases reference counter, deallocate the data when reference counter reaches 0
     CV_WRAP void release();
 
+    //! allocates or reuses underlying storage to fit the requested 2D size and type (no-op if already compatible)
+    void fit(int rows, int cols, int type);
+    void fit(Size size, int type);
+
     //! swaps with other smart pointer
     CV_WRAP void swap(GpuMat& mat);
 
@@ -394,7 +398,7 @@ struct CV_EXPORTS_W GpuData
 class CV_EXPORTS_W GpuMatND
 {
 public:
-    using SizeArray = std::vector<int>;
+    using SizeArray = MatShape;
     using StepArray = std::vector<size_t>;
     using IndexArray = std::vector<int>;
 
@@ -409,7 +413,7 @@ public:
     @param type Array type. Use CV_8UC1, ..., CV_16FC4 to create 1-4 channel matrices, or
     CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
     */
-    GpuMatND(SizeArray size, int type);
+    GpuMatND(const MatShape& shape, int type);
 
     /** @overload
     @param size Array of integers specifying an n-dimensional array shape.
@@ -424,7 +428,7 @@ public:
     (if specified, the last step must be equal to the element size, otherwise it will be added as such).
     If not specified, the matrix is assumed to be continuous.
     */
-    GpuMatND(SizeArray size, int type, void* data, StepArray step = StepArray());
+    GpuMatND(const MatShape& shape, int type, void* data, StepArray step = StepArray());
 
     /** @brief Allocates GPU memory.
     Suppose there is some GPU memory already allocated. In that case, this method may choose to reuse that
@@ -433,12 +437,19 @@ public:
     (i.e., isSubmatrix() is false). In other words, this method guarantees that the GPU memory allocated by
     this method is always continuous and is not a sub-region of another GpuMatND.
     */
-    void create(SizeArray size, int type);
+    void create(const MatShape& shape, int type);
 
     void release();
 
     void swap(GpuMatND& m) noexcept;
 
+    /** @brief Allocates or reuses underlying storage to fit the requested n-D size and type.
+    - No-op if already compatible (sufficient capacity, continuous, not a submatrix, not external, no ROI).
+    - Reallocates otherwise.
+    Mirrors 2D GpuMat::fit semantics for multi-dimensional tensors.
+    */
+    void fit(const MatShape& shape, int type);
+
     /** @brief Creates a full copy of the array and the underlying data.
     The method creates a full copy of the array. It mimics the behavior of Mat::clone(), i.e.
     the original step is not taken into account. So, the array copy is a continuous array
@@ -538,7 +549,7 @@ public:
 
 private:
     //! internal use
-    void setFields(SizeArray size, int type, StepArray step = StepArray());
+    void setFields(MatShape size, int type, StepArray step = StepArray());
 
 public:
     /*! includes several bit-fields:
@@ -553,7 +564,7 @@ public:
     int dims;
 
     //! shape of this array
-    SizeArray size;
+    MatShape size;
 
     /*! step values
     Their semantics is identical to the semantics of step for Mat.
@@ -580,6 +591,11 @@ private:
     size_t offset;
 };
 
+// Provide namespace-level aliases for nested array types used in bindings
+using SizeArray = GpuMatND::SizeArray;
+using StepArray = GpuMatND::StepArray;
+using IndexArray = GpuMatND::IndexArray;
+
 /** @brief Creates a continuous matrix.
 
 @param rows Row count.

modules/core/include/opencv2/core/cuda.inl.hpp

@@ -418,10 +418,10 @@ GpuMatND::GpuMatND() :
 }
 
 inline
-GpuMatND::GpuMatND(SizeArray _size, int _type) :
+GpuMatND::GpuMatND(const MatShape& _shape, int _type) :
     flags(0), dims(0), data(nullptr), offset(0)
 {
-    create(std::move(_size), _type);
+    create(_shape, _type);
 }
 
 inline