[Fix] Tensor::setDataFormat handling of DataFormat::Default

Context

Fix aidge_backend_cpu#42 (closed)

Implement back the behaviour of setDataFormat lost in !314 (merged) that transposing Tensor from/to DataFormat::Default does not lead to any data transposition. Only the dataFormat is changed.

Major modification

• upd: in DataFormat.hpp, change 'getDataFormatTranspose' name to 'getPermutationMapping'
• upd: use X-macro to keep DataFormat enum, string value, permutation with NCHW and description in synch
• add: number of dimensions information to X-macro
• upd: change 'DataFormatTransposeDict' array to 'permutationFromNCHW' and move it to a static function 'getPermutationFromNCHW'
• fix: do not transpose data in Tensor if current or new format is Default

include/aidge/data/DataFormat.hpp

@@ -17,63 +17,84 @@
 
 #include "aidge/utils/logger/EnumString.hpp"
 
+#define PERM(...) { __VA_ARGS__ }
+
+//   (EnumName, StringLiteral, NumDims, Permutation (0-indexed), Description)
+#define LIST_DATAFORMAT_ATTR(X)                                                      \
+    X(Default, "Default",         0, PERM(), "Unspecified format: interpretation is implementation-dependent"), \
+    X(Any,         "Any",         0, PERM(), "Any format is valid"),                    \
+    X(NCHW,        "NCHW",        4, PERM(0, 1, 2, 3), "4D format: [batch][channel][height][width]"), \
+    X(NHWC,        "NHWC",        4, PERM(0, 2, 3, 1), "4D format: [batch][height][width][channel]"), \
+    X(CHWN,        "CHWN",        4, PERM(1, 2, 3, 0), "4D format: [channel][height][width][batch]"), \
+    X(NCDHW,       "NCDHW",       5, PERM(0, 1, 2, 3, 4), "5D format: [batch][channel][depth][height][width]"), \
+    X(NDHWC,       "NDHWC",       5, PERM(0, 2, 3, 4, 1), "5D format: [batch][depth][height][width][channel]"), \
+    X(CDHWN,       "CDHWN",       5, PERM(1, 2, 3, 4, 0), "5D format: [channel][depth][height][width][batch]")
+
+#define NB_DFORMAT 8
+
 namespace Aidge {
 
 /**
- * @brief Enumeration of supported tensor data layouts
+ * @brief Enumeration of supported tensor data layouts.
  *
- * Represents different memory layout formats for multi-dimensional tensors:
+ * Represents different memory layouts for multi-dimensional tensors.
+ * The dimensions typically represent:
  * - N: Batch size
  * - C: Channels
  * - H: Height
  * - W: Width
  * - D: Depth (for 3D tensors)
+ *
+ * The enum values are generated via the X-macro.
  */
-enum class DataFormat {
-    Default,    ///< Default format, implementation dependent
-    NCHW,      ///< 4D format: [batch][channel][height][width]
-    NHWC,      ///< 4D format: [batch][height][width][channel]
-    CHWN,      ///< 4D format: [channel][height][width][batch]
-    NCDHW,     ///< 5D format: [batch][channel][depth][height][width]
-    NDHWC,     ///< 5D format: [batch][depth][height][width][channel]
-    CDHWN,     ///< 5D format: [channel][depth][height][width][batch]
-    Any        ///< Unspecified format
+enum class DataFormat : int {
+#define X(enumName, str, nb, arr, desc) enumName
+    LIST_DATAFORMAT_ATTR(X)
+#undef X
 };
 
-using DataFormatTranspose = std::array<std::size_t, 5>;
-
 /**
- * @brief Dictionary of transpose operations between different formats
+ * @brief Alias for a fixed-size array representing a permutation.
  *
- * Contains permutation arrays to convert between different data formats.
- * @warning In this array only, dimension index starts at 1
- * (0 is reserved as default value).
+ * A DataFormatTranspose is an array of 5 size_t values that represents a permutation mapping
+ * (0-indexed) for dimension reordering. Only the first NumDims elements are valid.
  */
-constexpr std::array<DataFormatTranspose, 7> DataFormatTransposeDict = {{
-    {},                 // Default
-    {1, 2, 3, 4},      // NCHW
-    {1, 3, 4, 2},      // NHWC
-    {2, 3, 4, 1},      // CHWN
-    {1, 2, 3, 4, 5},   // NCDHW
-    {1, 3, 4, 5, 2},   // NDHWC
-    {2, 3, 4, 5, 1}    // CDHWN
-}};
+using DataFormatTranspose = std::array<std::size_t, 5>;
 
 /**
- * @brief Get the permutation array for converting between data formats
+ * @brief Compute the permutation mapping to convert tensor data from a source format
+ *        to a destination format.
+ *
+ * This function performs the following steps:
+ *   1. If either format is DataFormat::Any, a runtime error is thrown.
+ *   2. If either format is DataFormat::Default or both dformat are equal, the identity permutation is returned.
+ *   3. Otherwise, the function retrieves the permutation mappings (from default NCHW)
+ *      for both the source and destination formats.
+ *   4. It computes the inverse of the source mapping (from source to default ordering).
+ *   5. Finally, it composes the inverse with the destination mapping to produce the
+ *      permutation mapping from source to destination.
+ *
+ * @param src The source data format.
+ * @param dst The destination data format.
+ * @return constexpr DataFormatTranspose The computed permutation array (0-indexed).
  *
- * @param src Source data format
- * @param dst Destination data format
- * @return DataFormatTranspose Permutation array to achieve the format conversion
+ * @throws std::runtime_error if either src or dst is DataFormat::Any.
+ * @pre The source and destination formats must have the same number of dimensions.
  */
-DataFormatTranspose getDataFormatTranspose(const DataFormat& src, const DataFormat& dst);
+DataFormatTranspose getPermutationMapping(const DataFormat& src, const DataFormat& dst);
 
 } // namespace Aidge
 
 namespace {
 template <>
-const char* const EnumStrings<Aidge::DataFormat>::data[]
-    = {"Default", "NCHW", "NHWC", "CHWN", "NCDHW", "NDHWC", "CDHWN", "Any"};
+struct EnumStrings<Aidge::DataFormat> {
+    static const char* const data[];
+};
+constexpr const char* const EnumStrings<Aidge::DataFormat>::data[] = {
+#define X(EnumName, Str, NumDims, Perm, Desc) Str
+    LIST_DATAFORMAT_ATTR(X)
+#undef X
+};
 }
 
 namespace Aidge {