3 // FIR FILTER COMPTUED DIRECTLY ON INPUT WITH NO
5 // TWO OUTPUTS PER ITERATION
6 // This program computes a FIR filter without maintaining a buffer of internal
8 // This example computes two output samples per inner loop. The following
9 // diagram shows the alignment required for signal x and coefficients c:
11 // c0 c1 c2 c3 c4 -> output z(0)=x0*c0 + x1*c1 + ...
12 // c0 c1 c2 c3 c4 -> z(1)=x1*c0 + x2*c1 + ...
15 // Z(k) = \ c(n) * x(n+k)
19 // Naive, first stab at spliting this for dual MACS.
22 // R(k) = \ (x(2n) * y(2n+k)) + \ (x(2n-1) * y(2n-1+k))
26 // Alternate, better partitioning for the machine.
29 // R(0) = \ x(n) * y(n)
35 // R(1) = \ x(n) * y(n+1)
41 // R(2) = \ x(n) * y(n+2)
47 // R(3) = \ x(n) * y(n+3)
55 // Okay in this verion the inner loop will compute R(2k) and R(2k+1) in parallel
58 // R(2k) = \ x(n) * y(n+2k)
64 // R(2k+1) = \ x(n) * y(n+2k+1)
70 // Sample pair x1 x0 is loaded into register R0, and coefficients c1 c0
71 // is loaded into register R1:
76 // | c1 c0 | compute two MACs: z(0)+=x0*c0, and z(1)+=x1*c0
78 // Now load x2 into lo half of R0, and compute the next two MACs:
83 // | c1 c0 | compute z(0)+=x1*c1 and z(1)+=x2*c1 (c0 not used)
85 // Meanwhile, load coefficient pair c3 c2 into R2, and x3 into hi half of R0:
90 // | c3 c2 | compute z(0)+=x2*c2 and z(1)+=x3*c2 (c3 not used)
92 // Load x4 into low half of R0:
97 // | c3 c2 | compute z(0)+=x3*c3 and z(1)+=x4*c3 (c2 not used)
99 // //This is a reference FIR function used to test: */
100 //void firf (float input[], float output[], float coeffs[],
101 // long input_size, long coeffs_size)
104 // for(i=0; i< input_size; i++){
106 // for(k=0; k < coeffs_size; k++)
107 // output[i] += input[k+i] * coeffs[k];
111 .include "testutils.inc"
115 R0 = 0; R1 = 0; R2 = 0;
116 P1 = 128 (X); // Load loop bounds in R5, R6, and divide by 2
119 // P0 holds pointer to input data in one memory
120 // bank. Increments by 2 after each inner-loop iter
123 // Pointer to coeffs in alternate memory bank.
126 // Pointer to outputs in any memory bank.
129 // Setup outer do-loop for M/2 iterations
130 // (2 outputs are computed per pass)
132 LSETUP ( L$0 , L$0end ) LC0 = P1 >> 1;
137 // Set-up inner do-loop for L/2 iterations
138 // (2 MACs are computed per pass)
140 LSETUP ( L$1 , L$1end ) LC1 = P2 >> 1;
142 // Load first two data elements in r0,
143 // and two coeffs into r1:
146 A1 = A0 = 0 || R0.H = W [ I0 ++ ] || R1 = [ I1 ++ ];
149 A1 += R0.H * R1.L, A0 += R0.L * R1.L || R0.L = W [ I0 ++ ] || NOP;
151 A1 += R0.L * R1.H, A0 += R0.H * R1.H || R0.H = W [ I0 ++ ] || R1 = [ I1 ++ ];
153 // Line 1: do 2 MACs and load next data element into RL0.
154 // Line 2: do 2 MACs, load next data element into RH0,
155 // and load next 2 coeffs
157 R0.H = A1, R0.L = A0;
159 // advance data pointer by 2 16b elements
163 [ I2 ++ ] = R0; // store 2 outputs
168 R0.L = W [ I2 ++ ]; DBGA ( R0.L , 0x0800 );
169 R0.L = W [ I2 ++ ]; DBGA ( R0.L , 0x1000 );
170 R0.L = W [ I2 ++ ]; DBGA ( R0.L , 0x2000 );
171 R0.L = W [ I2 ++ ]; DBGA ( R0.L , 0x1000 );
172 R0.L = W [ I2 ++ ]; DBGA ( R0.L , 0x0800 );
187 .space ((128-10)*2); // must pad with zeros or uninitialized values.
197 .space ((64-6)*2); // must pad with zeros or uninitialized values.