This project contains a header file that defines a N-dimensional array (variadic) template. This definition takes advantage of data locality which improves performance over e.g. nested std::vectors.
Mathematical expression are also optimized using lazy evaluation which is implemented with expression templates. Expression such as (A+B)*C is compiled to a single for loop containing (A[i]+B[i])*C[i], which makes it as efficient as hand-written C code. This also works when evaluating specific array element since only the specified element is effectively computed (e.g. (A+B)(i,j,k) simply compiles to (A(i,j,k)+B(i,j,k))).
The file speed_test.cpp is a demonstration of the speed difference between this template and nested std::valarray (which also uses lazy evaluation) as well as other interesting performance metrics. test.cpp contains tests but can also serve as a comprehensive list of features.
Note: Must be compiled with c++20
An array of double with shape [Dim_1,Dim_2,Dim_3] is defined with
Array<double, Dim_1,Dim_2,Dim_3> A;
By default, the values are not initialized. The array can be set to a specific value with
A = value;
This can be combined:
Array<double, Dim_1,Dim_2,Dim_3> B = value;
// OR
Array<double, Dim_1,Dim_2,Dim_3> C(value);
Arrays can be declared with other sub-arrays if dimensions match. In the example below, B contains Dim_1 copies of A.
Array<double, Dim_2,Dim_3> A = 42;
Array<double, Dim_1,Dim_2,Dim_3> B(A);
Finally, they can be created from expressions:
Array<double, Dim_1,Dim_2,Dim_3> B = 2*A - 1;
The usual mathematical, boolean and bitwise operations are implemented (in parallel if OpenMP is enabled) and use lazy evaluation. By default, all operations are performed element-wise, except in obvious situation such as for min/max or sum which all return a single value.
The type of the resulting array is automatically determined based on the operation and its operands.
For binary operation (e.g. + or <=), one of the operand can be a single value (e.g. a scalar) that is convertible to an element of the array.
C = 2*A + pow(A,B) - 2; // Element-wise pow(A,B): a^b
D = A<2; // Would evaluate to a boolean array
E = floor(1.5*A); // Would evaluate to an int array
Note: Using the auto keyword does not perform the computation and keeps an expression. To force the evaluation, use .eval()
auto C = A+B; // Array expression
auto D = A+B.eval(); // Array
There are 3 ways to access array elements
Array::operator() with the right amount of indices. This returns a reference to the corresponding element.
A(4,3,2) = 5;
std::cout<<A(4,3,2)<<std::endl;
5
operator[] with 1 index. This returns a slice of the original array at the specified index.
A[2] = 5;
std::cout<<A[2]<<std::endl;
5 5 5
5 5 5
5 5 5
Mask indexing. This can be used to accessed every elements respecting a specific condition:
A[B<2] = 42;
The size of each dimension can be accessed with Array::size(dim), where dim is the dimension (starts at 0). The default value for dim is 0 so A.size() returns the first dimension.
.all() can be used on boolean arrays to check if all elements are true.
As mentioned before, this requires c++20.
To get a useful speed-up from lazy evaluation, a reasonable optimization level is required (at least -O1 on gcc). The makefile uses more aggresive optimizations (-Ofast -march=native) which could cause some issues. Consider reducing -Ofast to -O2 (or -O3) if you encounter issues. Some tests may fail with -Ofast but should run fine with -O2 or even -O3. This comes from the flag -ffast-math and more precisely -funsafe-math-optimizations on gcc.
Operations are parallelised with OpenMP using multiple cores as well as simd operations.
Note that simd operations will probably be automatically performed even without OpenMP if you enable a high optimization level (at least -O3 on gcc).