xref: /llvm-project/clang/docs/MatrixTypes.rst (revision 6e58e99a07facd73547f7fa1a6f8c0719d8af30d)
1==================
2Matrix Types
3==================
4
5.. contents::
6   :local:
7
8.. _matrixtypes:
9
10Clang provides a C/C++ language extension that allows users to directly express
11fixed-size 2-dimensional matrices as language values and perform arithmetic on
12them.
13
14This feature is currently experimental, and both its design and its
15implementation are in flux.
16
17Draft Specification
18===================
19
20Matrix Type
21-----------
22
23A matrix type is a scalar type with an underlying *element type*, a constant
24number of *rows*, and a constant number of *columns*. Matrix types with the same
25element type, rows, and columns are the same type. A value of a matrix type
26includes storage for ``rows * columns`` values of the *element type*. The
27internal layout, overall size and alignment are implementation-defined.
28
29The maximum of the product of the number of rows and columns is
30implementation-defined. If that implementation-defined limit is exceeded, the
31program is ill-formed.
32
33Currently, the element type of a matrix is only permitted to be one of the
34following types:
35
36* an integer type (as in C23 6.2.5p22), but excluding enumerated types, ``bool``,
37  and ``_BitInt`` types whose width is not a power of 2;
38* the standard floating types ``float`` or ``double``;
39* a half-precision floating point type, if one is supported on the target.
40
41Other types may be supported in the future.
42
43Matrix Type Attribute
44---------------------
45
46Matrix types can be declared by adding the ``matrix_type`` attribute to the
47declaration of a *typedef* (or a C++ alias declaration). The underlying type
48of the *typedef* must be a valid matrix element type. The
49attribute takes two arguments, both of which must be integer constant
50expressions that evaluate to a value greater than zero. The first specifies the
51number of rows, and the second specifies the number of columns. The underlying
52type of the *typedef* becomes a matrix type with the given dimensions and an
53element type of the former underlying type.
54
55If a declaration of a *typedef-name* has a ``matrix_type`` attribute, then all
56declaration of that *typedef-name* shall have a matrix_type attribute with the
57same element type, number of rows, and number of columns.
58
59Standard Conversions
60--------------------
61
62The standard conversions are extended as follows. Note that these conversions
63are intentionally not listed as satisfying the constraints for assignment,
64which is to say, they are only permitted as explicit casts, not as implicit
65conversions.
66
67A value of matrix type can be converted to another matrix type if the number of
68rows and columns are the same and the value's elements can be converted to the
69element type of the result type. The result is a matrix where each element is
70the converted corresponding element.
71
72A value of any real type (as in C23 6.2.5p14) can be converted to a matrix type
73if it can be converted to the element type of the matrix. The result is a
74matrix where all elements are the converted original value.
75
76If the number of rows or columns differ between the original and resulting
77type, the program is ill-formed.
78
79
80Arithmetic Conversions
81----------------------
82
83The usual arithmetic conversions are extended as follows.
84
85Insert at the start:
86
87* If both operands are of matrix type, no arithmetic conversion is performed.
88* If one operand is of matrix type and the other operand is of a real type,
89  convert the real type operand to the matrix type
90  according to the standard conversion rules.
91
92Matrix Type Element Access Operator
93-----------------------------------
94
95An expression of the form ``E1 [E2] [E3]``, where ``E1`` has matrix type ``cv
96M``, is a matrix element access expression.  Let ``T`` be the element type
97of ``M``, and let ``R`` and ``C`` be the number of rows and columns in ``M``
98respectively.  The index expressions shall have integral or unscoped
99enumeration type and shall not be uses of the comma operator unless
100parenthesized.  The first index expression shall evaluate to a
101non-negative value less than ``R``, and the second index expression shall
102evaluate to a non-negative value less than ``C``, or else the expression has
103undefined behavior.  If ``E1`` is a prvalue, the result is a prvalue with type
104``T`` and is the value of the element at the given row and column in the matrix.
105Otherwise, the result is a glvalue with type ``cv T`` and with the same value
106category as ``E1`` which refers to the element at the given row and column in
107the matrix.
108
109Programs containing a single subscript expression into a matrix are ill-formed.
110
111**Note**: We considered providing an expression of the form
112``postfix-expression [expression]`` to access columns of a matrix. We think
113that such an expression would be problematic once both column and row major
114matrixes are supported: depending on the memory layout, either accessing columns
115or rows can be done efficiently, but not both. Instead, we propose to provide
116builtins to extract rows and columns from a matrix. This makes the operations
117more explicit.
118
119Matrix Type Binary Operators
120----------------------------
121
122Given two matrixes, the ``+`` and ``-`` operators perform element-wise addition
123and subtraction, while the ``*`` operator performs matrix multiplication.
124``+``, ``-``, ``*``, and ``/`` can also be used with a matrix and a scalar
125value, applying the operation to each element of the matrix.
126
127Earlier versions of this extension did not support division by a scalar.
128You can test for the availability of this feature with
129``__has_extension(matrix_types_scalar_division)``.
130
131For the expression ``M1 BIN_OP M2`` where
132
133* ``BIN_OP`` is one of ``+`` or ``-``, one of ``M1`` and ``M2`` is of matrix
134  type, and the other is of matrix type or real type; or
135* ``BIN_OP`` is ``*``, one of ``M1`` and ``M2`` is of matrix type, and the
136   other is of a real type; or
137* ``BIN_OP`` is ``/``, ``M1`` is of matrix type, and ``M2`` is of a real type:
138
139* The usual arithmetic conversions are applied to ``M1`` and ``M2``. [ Note: if ``M1`` or
140  ``M2`` are of a real type, they are broadcast to matrices here. — end note ]
141* ``M1`` and ``M2`` shall be of the same matrix type.
142* The result is equivalent to Res in the following where col is the number of
143  columns and row is the number of rows in the matrix type:
144
145.. code-block:: c++
146
147  decltype(M1) Res;
148  for (int C = 0; C < col; ++C)
149    for (int R = 0; R < row; ++R)
150      Res[R][C] = M1[R][C] BIN_OP M2[R][C];
151
152Given the expression ``M1 * M2`` where ``M1`` and ``M2`` are of matrix type:
153
154* The usual arithmetic conversions are applied to ``M1`` and ``M2``.
155* The type of ``M1`` shall have the same number of columns as the type of ``M2`` has
156  rows. The element types of ``M1`` and ``M2`` shall be the same type.
157* The resulting type, ``MTy``, is a matrix type with the common element type,
158  the number of rows of ``M1`` and the number of columns of ``M2``.
159* The result is equivalent to ``Res`` in the following where ``EltTy`` is the
160  element type of ``MTy``, ``col`` is the number of columns, ``row`` is the
161  number of rows in ``MTy`` and ``inner`` is the number of columns of ``M1``:
162
163.. code-block:: c++
164
165  MTy Res;
166  for (int C = 0; C < col; ++C) {
167    for (int R = 0; R < row; ++R) {
168      EltTy Elt = 0;
169      for (int K = 0; K < inner; ++K) {
170        Elt += M1[R][K] * M2[K][C];
171    }
172    Res[R][C] = Elt;
173  }
174
175All operations on matrix types match the behavior of the element type with
176respect to signed overflows.
177
178With respect to floating-point contraction, rounding and environment rules,
179operations on matrix types match the behavior of the elementwise operations
180in the corresponding expansions provided above.
181
182Operations on floating-point matrices have the same rounding and floating-point
183environment behavior as ordinary floating-point operations in the expression's
184context. For the purposes of floating-point contraction, all calculations done
185as part of a matrix operation are considered intermediate operations, and their
186results need not be rounded to the format of the element type until the final
187result in the containing expression. This is subject to the normal restrictions
188on contraction, such as ``#pragma STDC FP_CONTRACT``.
189
190For the ``+=``, ``-=`` and ``*=`` operators the semantics match their expanded
191variants.
192
193Matrix Type Builtin Operations
194------------------------------
195
196Each matrix type supports a collection of builtin expressions that look like
197function calls but do not form an overload set. Here they are described as
198function declarations with rules for how to construct the argument list types
199and return type and the library description elements from
200[library.description.structure.specifications]/3 in the C++ standard.
201
202Definitions:
203
204* *M*, *M1*, *M2*, *M3* - Matrix types
205* *T* - Element type
206* *row*, *col* - Row and column arguments respectively.
207
208
209``M2 __builtin_matrix_transpose(M1 matrix)``
210
211**Remarks**: The return type is a cv-unqualified matrix type that has the same
212element type as ``M1`` and has the same number of rows as ``M1`` has columns and
213the same number of columns as ``M1`` has rows.
214
215**Returns**: A matrix ``Res`` equivalent to the code below, where ``col`` refers to the
216number of columns of ``M``, and ``row`` to the number of rows of ``M``.
217
218**Effects**: Equivalent to:
219
220.. code-block:: c++
221
222  M Res;
223  for (int C = 0; C < col; ++C)
224    for (int R = 0; R < row; ++R)
225      Res[C][R] = matrix[R][C];
226
227
228``M __builtin_matrix_column_major_load(T *ptr, size_t row, size_t col, size_t columnStride)``
229
230**Mandates**: ``row`` and ``col`` shall be integral constants greater than 0.
231
232**Preconditions**: ``columnStride`` is greater than or equal to ``row``.
233
234**Remarks**: The return type is a cv-unqualified matrix type with an element
235type of the cv-unqualified version of ``T`` and a number of rows and columns equal
236to ``row`` and ``col`` respectively. The parameter ``columnStride`` is optional
237and if omitted ``row`` is used as ``columnStride``.
238
239**Returns**: A matrix ``Res`` equivalent to:
240
241.. code-block:: c++
242
243  M Res;
244  for (size_t C = 0; C < col; ++C) {
245    for (size_t R = 0; R < row; ++K)
246      Res[R][C] = ptr[R];
247    ptr += columnStride
248  }
249
250
251``void __builtin_matrix_column_major_store(M matrix, T *ptr, size_t columnStride)``
252
253**Preconditions**: ``columnStride`` is greater than or equal to the number of rows in ``M``.
254
255**Remarks**: The type ``T`` is the const-unqualified version of the matrix
256argument’s element type. The parameter ``columnStride`` is optional and if
257omitted, the number of rows of ``M`` is used as ``columnStride``.
258
259**Effects**: Equivalent to:
260
261.. code-block:: c++
262
263  for (size_t C = 0; C < columns in M; ++C) {
264    for (size_t R = 0; R < rows in M; ++K)
265      ptr[R] = matrix[R][C];
266    ptr += columnStride
267  }
268
269
270TODOs
271-----
272
273TODO: Does it make sense to allow M::element_type, M::rows, and M::columns
274where M is a matrix type? We don’t support this anywhere else, but it’s
275convenient. The alternative is using template deduction to extract this
276information. Also add spelling for C.
277
278Future Work: Initialization syntax.
279
280
281Decisions for the Implementation in Clang
282=========================================
283
284This section details decisions taken for the implementation in Clang and is not
285part of the draft specification.
286
287The elements of a  value of a matrix type are laid out in column-major order
288without padding.
289
290We propose to provide a Clang option to override this behavior and allow
291contraction of those operations (e.g. *-ffp-contract=matrix*).
292
293TODO: Specify how matrix values are passed to functions.
294