Материал: spra118

SPRA118

NOTE: You load F[|I|] << 11 to scale it with DML (Q11 format).

5.4.18TRANS

Function: Transition detector. tr(k) = 1 td(k–1) = 1 and |dq(k)| u 24 * 2yl(k–1)

Inputs: TD, YL (high word), DQ

Output: branch to UPA2 () (TR = 0), branch to TRIGB (TR = 1)

Cycles: Min: 6 (usual), Max: 25

NOTE: At the end of this routine, either perform TRIGA and TRIGB, or continue the program normally. Note that G.726 recommendation indicates in the text td(k) instead of td(k–1), and yl(k) instead of yl(k–1), but it is in contradiction with the further block description.

5.4.19TRIGA

Function: Speed control trigger block. ap(k) = ap(k–1) if tr(k) = 0, ap(k) = 1 if tr(k) = 1 Inputs: B (B u 0 tr = 1)

Outputs: AP

Cycles: Min: 0 (usual), Max: 2

NOTE: By using long-word instruction, td(k) is initialized at the same time as ap(k), performing by this way a part of TRIGB (20).

5.4.20TRIGB

Function: Predictor trigger block. If tr(k) = 1, ai(k) = 0 for i = 1, 2; bi(k) = 0 for i = 1, ..., 6; and td(k) = 0

Input: none

Output: A1, A2, B1, ..., B6, TD, and branch to FUNCTW() if performed

Cycles: Min: 0 (usual), Max: 7

NOTE: This routine (just as TRIGA) is executed in function of the precedent “conditional branch” at the end of the routine, TRANS (18). See TRIGA (19) for execution conditions. If tr(k) = 1 (transition is detected), TRIGB is performed: A1, A2, B1, ..., B6 are set to 0, then routines (21) to (29) are skipped to avoid Ai and Bi adaptation. Otherwise, (transition not detected), TRIGB is not performed: Ai, Bi, and TD keep their value, and are then adapted (for Ai, Bi: routines (21–23) and (25–26)), or calculated (for TD). Note that TD is initialized at the same time as AP (see TRIGA (19)).

5.4.21UPA2

Function: Update a2 coefficient of second order predictor. a2(k) = (1 – 2–7)a2(k–1) + 2–7{sgn[p(k)] sgn[p(k–2)] – f[a1(k–1)] sgn[p(k)] sgn[p(k–1)]}

Inputs: PK0high (PK0 = sgn(p(k)), PK0low (P(k)), PK1, PK2, A1, A2

SPRA118

Output: B (Ui)

Cycles: Min: 0, Max: 12 (usual)

NOTE: DQi sign is pointed by AR7 in table DQFLOAT. This routine is partially executed before each UPBi (26). This routine is not performed if tr(k) = 1.

5.4.26UPB

Function: Update for coefficients of sixth-order predictor. bi(k) = (1–2–8)bi(k–1) + 2–7 Ui(k) for 16-, 24-, 32-Kbps coding; bi(k) = (1–2–9)bi(k–1) + 2–7 Ui(k) for 40 Kbps coding

Inputs: B (Ui), Bi, DQ, SHIFT

Outputs: B1 (B1P), ..., B6 (B6P)

Cycles: Min: 0, Max: 40 (usual)

NOTE: If DQ = 0, then Ui is forced to be 0. SHIFT is –8 for 16, 24, 32 Kbps, and is –9 for 40 Kbps coding. It corresponds to the term 2–8 or 2–9 for Bi adaptation. This routine is not performed if tr(k) = 1.

5.4.27TONE

Function: Partial band-signal detection. td(k) = 1 if a2(k) t –0.71875, td(k) = 0. Otherwise:

Input: A2 (A2P)

Output: TD (TDP)

Cycles: Min: 0, Max: 3

NOTE: Computing value for –0.71875 is –11776 (Q14 format). This routine is not performed if tr(k) = 1.

5.4.28SUBTC

Function: Compute magnitude of the difference of shortand long-term functions of quantizer output sequence, and then perform threshold comparison for quantizing speed control

parameter. Ax = 0 if y(k) w 3 and |dms(k–1) – dml(k–1)| w 2–3 dml(k–1) and td(k–1) = 0, Ax = 1. Otherwise:

Inputs: DMS, DML, TD (TDP), Y

Output: A (AX << 9)

Cycles: Min: 0 (rare), Med: 6, Max: 21

NOTE: If td(k) = 1 (TD = –1), the routine is limited to 6 execution cycles, else if |dms(k–1) – dml(k–1)| w 2–3 dml(k–1), the routine is limited to 19 cycles. Otherwise, the cycle number is

21.If tr(k) = 1, the routine is not performed.

5.4.29FILTC

Function: Low-pass filter of speed control parameter. ap(k) = (1– 2–4)ap(k–1) + 2–3 Ax

SPRA118

Cycles: 5

NOTE: YL must be calculated before applying the 2–6 factor. As the theoretical format of YL is Q25, (–YL) is truncated to obtain Q15 format, as specified in G.726 recommendation.

5.4.34FLOATA

Function: Convert 16-bit 2’s complement to floating-point. DQ → (DQ0 mantissa, DQ0 exponent, DQ0 sign)

Input: DQ, SIGN (DQS), AR7 (points to DQ6 exponent)

Output: DQ0 exponent, DQ0 mantissa, DQ0 sign (in DQFLOAT buffer), AR7 (points to DQ5 exponent)

Cycles: 13

NOTE: Exponent EXP of DQ is defined by: 2EXP–1 v |DQ| 2EXP . Mantissa MANT of DQ is obtained with the 6 most significant bits (MSB) of DQ. MANT has the following limits: 0 v MANT t 1, which implies that MANT format is Q6. Sign of DQ is 0 if DQ positive, –1 if DQ negative. If DQ 0 0, find the exponent of DQ by means of the EXP and DSUBT instructions, and MANT via the NORM instruction. If DQ = 0, EXP is 0 and MANT is defined to be 1/2 (= 32).For this particular case, the sign is given by the variable SIGN (DQS = sign of DQ given by the inverse quantizer). Note that DQ = 0 does not imply that DQS = 0.

5.4.35FLOATB

Function: Convert 16-bit 2’s complement to floating point. SR → (SR0 mantissa, SR0 exponent, SR0 sign)

Input: SD (SR), AR6 (points to SR2 sign)

Output: SR0, AR5 (points to SR1 exponent)

Cycles: 14

NOTE: See FLOATA for definitions of exponent, mantissa, and sign; but contrary to DQ, SR = 0 always implies that the sign of SR is also 0 (this means that, in this case, SR is positive).

5.4.36DELAY

Function: Memory block. For the input x, the output is given by y(k) = x(k–1).

Input: x

Output: y

Cycles: 10

NOTE: This routine applies for all delayed variables. For the one-time delayed variables such as DMS, DML, AP, Ai, Bi, TD, YL, YU, x(k + 1) has the same memory location as x(k). So, the delay was applied at the time these variables were updated. Thus, this routine really applies to the variables that are delayed several times, such as Pki, SRi, DQi. As for PKi, PK2 location just follows the PK1 location, so the instruction DELAY automatically realizes the PK2 update. In the case of Sri and Sri, these variables are automatically delayed, due to the use of two circular buffers (DQFLOAT for DQi and SRFLOAT for SRi). Only the addresses of the next DQ6 and the next SR2 must be saved (ADDQ6 and ADSR2).