Alexandre Lision | 8af73cb | 2013-12-10 14:11:20 -0500 | [diff] [blame] | 1 | /* |
| 2 | * The configuration constants below govern |
| 3 | * the number of bits in the input sample and filter coefficients, the |
| 4 | * number of bits to the right of the binary-point for fixed-point math, etc. |
| 5 | * |
| 6 | */ |
| 7 | |
| 8 | /* Conversion constants */ |
| 9 | #define Nhc 8 |
| 10 | #define Na 7 |
| 11 | #define Np (Nhc+Na) |
| 12 | #define Npc (1<<Nhc) |
| 13 | #define Amask ((1<<Na)-1) |
| 14 | #define Pmask ((1<<Np)-1) |
| 15 | #define Nh 16 |
| 16 | #define Nb 16 |
| 17 | #define Nhxn 14 |
| 18 | #define Nhg (Nh-Nhxn) |
| 19 | #define NLpScl 13 |
| 20 | |
| 21 | /* Description of constants: |
| 22 | * |
| 23 | * Npc - is the number of look-up values available for the lowpass filter |
| 24 | * between the beginning of its impulse response and the "cutoff time" |
| 25 | * of the filter. The cutoff time is defined as the reciprocal of the |
| 26 | * lowpass-filter cut off frequence in Hz. For example, if the |
| 27 | * lowpass filter were a sinc function, Npc would be the index of the |
| 28 | * impulse-response lookup-table corresponding to the first zero- |
| 29 | * crossing of the sinc function. (The inverse first zero-crossing |
| 30 | * time of a sinc function equals its nominal cutoff frequency in Hz.) |
| 31 | * Npc must be a power of 2 due to the details of the current |
| 32 | * implementation. The default value of 512 is sufficiently high that |
| 33 | * using linear interpolation to fill in between the table entries |
| 34 | * gives approximately 16-bit accuracy in filter coefficients. |
| 35 | * |
| 36 | * Nhc - is log base 2 of Npc. |
| 37 | * |
| 38 | * Na - is the number of bits devoted to linear interpolation of the |
| 39 | * filter coefficients. |
| 40 | * |
| 41 | * Np - is Na + Nhc, the number of bits to the right of the binary point |
| 42 | * in the integer "time" variable. To the left of the point, it indexes |
| 43 | * the input array (X), and to the right, it is interpreted as a number |
| 44 | * between 0 and 1 sample of the input X. Np must be less than 16 in |
| 45 | * this implementation. |
| 46 | * |
| 47 | * Nh - is the number of bits in the filter coefficients. The sum of Nh and |
| 48 | * the number of bits in the input data (typically 16) cannot exceed 32. |
| 49 | * Thus Nh should be 16. The largest filter coefficient should nearly |
| 50 | * fill 16 bits (32767). |
| 51 | * |
| 52 | * Nb - is the number of bits in the input data. The sum of Nb and Nh cannot |
| 53 | * exceed 32. |
| 54 | * |
| 55 | * Nhxn - is the number of bits to right shift after multiplying each input |
| 56 | * sample times a filter coefficient. It can be as great as Nh and as |
| 57 | * small as 0. Nhxn = Nh-2 gives 2 guard bits in the multiply-add |
| 58 | * accumulation. If Nhxn=0, the accumulation will soon overflow 32 bits. |
| 59 | * |
| 60 | * Nhg - is the number of guard bits in mpy-add accumulation (equal to Nh-Nhxn) |
| 61 | * |
| 62 | * NLpScl - is the number of bits allocated to the unity-gain normalization |
| 63 | * factor. The output of the lowpass filter is multiplied by LpScl and |
| 64 | * then right-shifted NLpScl bits. To avoid overflow, we must have |
| 65 | * Nb+Nhg+NLpScl < 32. |
| 66 | */ |
| 67 | |