The difference between the single-kernel case and this case is that the tap array contains eight elements and that the delay line is only 16 element deep. In the previous section, you saw that there were only two mul4 and mac4 intrinsics for cint16 x cint16 operations.
More interestingly is the one that uses a v16cint16 for the data register. Filter output compute for an eight tap filter on four lanes in parallel requires (8+3 = 11) data in the buffer. The delay-line should contain at least eight samples as the seven previous samples are needed for the computation of the first output.
The following image gives an idea of data update scheduling and how it is interleaved with mul4/mac4 operations (only the first eight outputs).
The related C++ code is as follows:
void SingleStream::FIR_MultiKernel_cout<NSamples,ShiftAcc>::filter(input_stream_cint16* sin,output_stream_cacc48* cout)
{
v8cint16 taps = *(v8cint16*) weights;
v16cint16 *ptr_delay_line = (v16cint16 *)delay_line;
v16cint16 data = *ptr_delay_line;
v4cacc48 acc = undef_v4cacc48();
// Computes 16 samples per iteration
for(int i=0;i<NSamples/16;i++)
chess_prepare_for_pipelining
chess_loop_range(NSamples/16,NSamples/16)
{
acc = mul4(data,1,0x3210,1,taps,0,0x0000,1);
acc = mac4(acc,data,3,0x3210,1,taps,2,0x0000,1);
data = upd_v(data, 2, readincr_v4(sin));
acc = mac4(acc,data,5,0x3210,1,taps,4,0x0000,1);
acc = mac4(acc,data,7,0x3210,1,taps,6,0x0000,1);
writeincr_v4(cout,acc);
acc = mul4(data,5,0x3210,1,taps,0,0x0000,1);
acc = mac4(acc,data,7,0x3210,1,taps,2,0x0000,1);
data = upd_v(data, 3, readincr_v4(sin));
acc = mac4(acc,data,9,0x3210,1,taps,4,0x0000,1);
acc = mac4(acc,data,11,0x3210,1,taps,6,0x0000,1);
writeincr_v4(cout,acc);
acc = mul4(data,9,0x3210,1,taps,0,0x0000,1);
acc = mac4(acc,data,11,0x3210,1,taps,2,0x0000,1);
data = upd_v(data, 0, readincr_v4(sin));
acc = mac4(acc,data,13,0x3210,1,taps,4,0x0000,1);
acc = mac4(acc,data,15,0x3210,1,taps,6,0x0000,1);
writeincr_v4(cout,acc);
acc = mul4(data,13,0x3210,1,taps,0,0x0000,1);
acc = mac4(acc,data,15,0x3210,1,taps,2,0x0000,1);
data = upd_v(data, 1, readincr_v4(sin));
acc = mac4(acc,data,1,0x3210,1,taps,4,0x0000,1);
acc = mac4(acc,data,3,0x3210,1,taps,6,0x0000,1);
writeincr_v4(cout,acc);
}
*ptr_delay_line = data;
}
This code is the one of FIR_MultiKernel_cout, the output is sent to the cascade stream using the writeincr_v4(cout,acc) instruction.
At the graph level, all kernels are first declared in a class:
class FIRGraph_4Kernels: public adf::graph
{
private:
kernel k[4];
public:
input_port in[4];
output_port out;
The constructor takes charge of the next operations. The first operation is to create the kernels: 1xFIR_MultiKernel_cout, 2xFIR_MultiKernel_cincout, 1xFIR_MultiKernel_cin
FIRGraph_4Kernels()
{
k[0] = kernel::create_object<SingleStream::FIR_MultiKernel_cout<NUM_SAMPLES,SHIFT>>(taps4_0);
k[1] = kernel::create_object<SingleStream::FIR_MultiKernel_cincout<NUM_SAMPLES,SHIFT>>(taps4_1);
k[2] = kernel::create_object<SingleStream::FIR_MultiKernel_cincout<NUM_SAMPLES,SHIFT>>(taps4_2);
k[3] = kernel::create_object<SingleStream::FIR_MultiKernel_cin<NUM_SAMPLES,SHIFT>>(taps4_3);
The AI Engine compiler needs to know the location of the source code for the kernels:
const int NChunks = 4;
for(int i=0;i<NChunks;i++)
{
runtime<ratio>(k[i]) = 0.9;
source(k[i]) = "aie_kernels/FirSingleStream.cpp";
headers(k[i]) = {"aie_kernels/FirSingleStream.h"};
}
To shorten the place time by a few seconds, constrain the core location. A single one is necessary because all the others are constrained by the cascade connection:
// Constraints: location of the first kernel in the cascade
location<kernel>(k[0]) = tile(25,0);
All the kernels need to discard a specific number of elements. This is handled by the initialization function as this must be done beforehand and only once:
// Discard first elements of the stream, depending on position in the cascade
initialization_function(k[0]) = "SingleStream::FIRinit<0>";
initialization_function(k[1]) = "SingleStream::FIRinit<8>";
initialization_function(k[2]) = "SingleStream::FIRinit<16>";
initialization_function(k[3]) = "SingleStream::FIRinit<24>";
Finally, all kernels must be connected together. This is done at the end of the constructor of the class:
// Cascade Connections and output connection
for(int i=0;i<NChunks-1;i++)
connect<cascade> (k[i].out[0],k[i+1].in[1]);
connect<stream> (k[NChunks-1].out[0],out);
// Input Streams connections
for(int i=0;i<NChunks;i++)
connect<stream>(in[i],k[i].in[0]);
The initialization function is simple. It simply reads data from the input stream. Because there is no argument, the raw API for stream access must be used:
template<int Delay>
void SingleStream::FIRinit()
{
for (int i = 0; i < Delay; ++i)
{
get_ss(0);
}
}