Matrix Multiplication with C++ AMP

Fri, September 9, 2011, 08:22 PM under GPGPU | ParallelComputing

As part of our API tour of C++ AMP, we looked recently at parallel_for_each. I ended that post by saying we would revisit parallel_for_each after introducing array and array_view. Now is the time, so this is part 2 of parallel_for_each, and also a post that brings together everything we've seen until now.

The code for serial and accelerated

Consider a naïve (or brute force) serial implementation of matrix multiplication 

0:  void MatrixMultiplySerial(std::vector<float>& vC, 
        const std::vector<float>& vA, 
        const std::vector<float>& vB, int M, int N, int W)
1:  {
2:    for (int row = 0; row < M; row++) 
3:    {
4:      for (int col = 0; col < N; col++)
5:      {
6:        float sum = 0.0f;
7:        for(int i = 0; i < W; i++)
8:          sum += vA[row * W + i] * vB[i * N + col];
9:        vC[row * N + col] = sum;
10:     }
11:   }
12: }

We notice that each loop iteration is independent from each other and so can be parallelized. If in addition we have really large amounts of data, then this is a good candidate to offload to an accelerator. First, I'll just show you an example of what that code may look like with C++ AMP, and then we'll analyze it. It is assumed that you included at the top of your file #include <amp.h>

13:  void MatrixMultiplySimple(std::vector<float>& vC, 
         const std::vector<float>& vA, 
         const std::vector<float>& vB, int M, int N, int W)
14:  {
15:    concurrency::array_view<const float,2> a(M, W, vA);
16:    concurrency::array_view<const float,2> b(W, N, vB);
17:    concurrency::array_view<float,2> c(M, N, vC); c.discard_data();
18:    concurrency::parallel_for_each(c.extent, 
19:    [=](concurrency::index<2> idx) restrict(amp) {
20:      int row = idx[0]; int col = idx[1];
21:      float sum = 0.0f;
22:      for(int i = 0; i < W; i++)
23:        sum += a(row, i) * b(i, col);
24:      c[idx] = sum;
25:    });
26:  }

First a visual comparison, just for fun: The beginning and end is the same, i.e. lines 0,1,12 are identical to lines 13,14,26. The double nested loop (lines 2,3,4,5 and 10,11) has beenMatrix Multiplication image from wikipedia transformed into a parallel_for_each call (18,19,20 and 25). The core algorithm (lines 6,7,8,9) is essentially the same (lines 21,22,23,24). We have extra lines in the C++ AMP version (15,16,17). Now let's dig in deeper.

Using array_view and extent

When we decided to convert this function to run on an accelerator, we knew we couldn't use the std::vector objects in the restrict(amp) function. So we had a choice of copying the data to the the concurrency::array<T,N> object, or wrapping the vector container (and hence its data) with a concurrency::array_view<T,N> object from amp.h – here we used the latter (lines 15,16,17). Now we can access the same data through the array_view objects (a and b) instead of the vector objects (vA and vB), and the added benefit is that we can capture the array_view objects in the lambda (lines 19-25) that we pass to the parallel_for_each call (line 18) and the data will get copied on demand for us to the accelerator.

Note that line 15 (and ditto for 16 and 17) could have been written as two lines instead of one:

  extent<2> e(M, W);
  array_view<const float, 2> a(e, vA);

In other words, we could have explicitly created the extent object instead of letting the array_view create it for us under the covers through the constructor overload we chose. The benefit of the extent object in this instance is that we can express that the data is indeed two dimensional, i.e a matrix. When we were using a vector object we could not do that, and instead we had to track via additional unrelated variables the dimensions of the matrix (i.e. with the integers M and W) – aren't you loving C++ AMP already?

Note that the const before the float when creating a and b, will result in the underling data only being copied to the accelerator and not be copied back – a nice optimization. A similar thing is happening on line 17 when creating array_view c, where we have indicated that we do not need to copy the data to the accelerator, through the discard_data call.

The kernel dispatch

On line 18 we make the call to the C++ AMP entry point (parallel_for_each) to invoke our parallel loop or, as some may say, dispatch our kernel.

The first argument we need to pass describes how many threads we want for this computation. For this algorithm we decided that we want exactly the same number of threads as the number of elements in the output matrix, i.e. in array_view c which will eventually update the vector vC. So each thread will compute exactly one result. Since the elements in c are organized in a 2-dimensional manner we can organize our threads in a two-dimensional manner too. We don't have to think too much about how to create the first argument (a extent) since the array_view object helpfully exposes that as a property. Note that instead of c.extent we could have written extent<2>(M, N) – the result is the same in that we have specified M*N threads to execute our lambda.

The second argument is a restrict(amp) lambda that accepts an index object. Since we elected to use a two-dimensional extent as the first argument of parallel_for_each, the index will also be two-dimensional and as covered in the previous posts it represents the thread ID, which in our case maps perfectly to the index of each element in the resulting array_view.

The kernel itself

The lambda body (lines 20-24), or as some may say, the kernel, is the code that will actually execute on the accelerator. It will be called by M*N threads and we can use those threads to index into the two input array_views (a,b) and write results into the output array_view ( c ).

The four lines (21-24) are essentially identical to the four lines of the serial algorithm (6-9). The only difference is how we index into a,b,c versus how we index into vA,vB,vC. The code we wrote with C++ AMP is much nicer in its indexing, because the dimensionality is a first class concept, so you don't have to do funny arithmetic calculating the index of where the next row starts, which you have to do when working with vectors directly (since they store all the data in a flat manner).

I skipped over describing line 20. Note that we didn't really need to read the two components of the index into temporary local variables. This mostly reflects my personal choice, in some algorithms to break down the index into local variables with names that make sense for the algorithm, i.e. in this case row and col. In other cases it may i,j,k or x,y,z, or M,N or whatever. Also note that we could have written line 24 as: c(idx[0], idx[1])=sum  or  c(row, col)=sum instead of the simpler c[idx]=sum

Targeting a specific accelerator

Imagine that we had more than one hardware accelerator on a system and we wanted to pick a specific one to execute this parallel loop on. So there would be some code like this anywhere before line 18:

  vector<accelerator> accs = MyFunctionThatChoosesSuitableAccelerators();
  accelerator acc = accs[0];

…and then we would modify line 18 so we would be calling another overload of parallel_for_each that accepts an accelerator_view as the first argument, so it would become:

  concurrency::parallel_for_each(acc.default_view, c.extent,

...and the rest of your code remains the same… how simple is that?