Introduction

Dot product example

If you want to compute the dot product:

\[ <scale*(\overrightarrow{a}+\overrightarrow{b}),\overrightarrow{c}*\overrightarrow{d}> \]

with CMSIS-DSP, you would write:

arm_add_f32(a,b,tmp1,NB);
arm_scale_f32(tmp1,scale,tmp2,NB);
arm_mult_f32(c,d,tmp3,NB);
arm_dot_prod_f32(tmp2,tmp3,NB,&r);

There are several limitations with this way of writing the code:

1. The code needs to be rewritten and the _f32 suffix changed if the developer wants to use another datatype
2. Temporary buffers need to be allocated and managed (tmp1,tmp2,tmp3,tmp4)
3. The four function calls are four different loops. It is not good for data locality and caches. The computation is not done in one pass
4. Each loop contains a small number of instructions. For instance, for the arm_add_f32, two loads, an add instruction and a store. It is not enough to enable the compiler to reorder the instructions to improve the performance

With this new C++ template library, you can write:

r = dot(scale*(a+b),c*d);

The code generated by this line computes the dot product in one pass with all the operators (+, *) included in the loop. There is no more any temporary buffers.

Vector operations

Let's look at another example:

\[ \overrightarrow{d} = \overrightarrow{a} + \overrightarrow{b} * \overrightarrow{c} \]

With the C++ library, it can be written as:

Vector<float32_t,NB> d = a + b * c;

Here again : all the vector operations (+,*) are done in one pass with one loop. There is no more any temporary buffer.

If you're coming from C and does not know anything about C++ templates, we have a very quick introduction : The minimum you need to know about C++ template to use this library.

You can also jump directly to an example with vector operations.