#include <ao_fec.h>
#include <stdio.h>
-#ifdef MEGAMETRUM
+#ifdef TELEMEGA
#include <ao.h>
#endif
}
#define NUM_STATE 8
-#define NUM_HIST 8
+#define NUM_HIST 24
+
+typedef uint32_t bits_t;
#define V_0 0xff
#define V_1 0x00
*/
uint8_t
-ao_fec_decode(const uint8_t *in, uint16_t len, uint8_t *out, uint8_t out_len, uint16_t (*callback)())
+ao_fec_decode(const uint8_t *in, uint16_t len, uint8_t *out, uint8_t out_len, uint16_t (*callback)(void))
{
static uint32_t cost[2][NUM_STATE]; /* path cost */
- static uint16_t bits[2][NUM_STATE]; /* save bits to quickly output them */
+ static bits_t bits[2][NUM_STATE]; /* save bits to quickly output them */
uint16_t i; /* input byte index */
uint16_t b; /* encoded symbol index (bytes/2) */
avail -= 2;
- /* Reset next costs to 'impossibly high' values so that
- * the first path through this state is cheaper than this
+ /* Compute path costs and accumulate output bit path
+ * for each state and encoded bit value. Unrolling
+ * this loop is worth about > 30% performance boost.
+ * Decoding 76-byte remote access packets is reduced
+ * from 14.700ms to 9.3ms. Redoing the loop to
+ * directly compare the two pasts for each future state
+ * reduces this down to 5.7ms
*/
- for (state = 0; state < NUM_STATE; state++)
- cost[n][state] = 0x7fffffff;
- /* Compute path costs and accumulate output bit path
- * for each state and encoded bit value
+ /* Ok, of course this is tricky, it's optimized.
+ *
+ * First, it's important to realize that we have 8
+ * states representing the combinations of the three
+ * most recent bits from the encoder. Flipping any
+ * of these three bits flips both output bits.
+ *
+ * 'state<<1' represents the target state for a new
+ * bit value of 0. '(state<<1)+1' represents the
+ * target state for a new bit value of 1.
+ *
+ * 'state' is the previous state with an oldest bit
+ * value of 0. 'state + 4' is the previous state with
+ * an oldest bit value of 1. These two states will
+ * either lead to 'state<<1' or '(state<<1)+1', depending
+ * on whether the next encoded bit was a zero or a one.
+ *
+ * m0 and m1 are the cost of coming to 'state<<1' from
+ * one of the two possible previous states 'state' and
+ * 'state + 4'.
+ *
+ * Because we know the expected values of each
+ * received bit are flipped between these two previous
+ * states:
+ *
+ * bitcost(state+4) = 510 - bitcost(state)
+ *
+ * With those two total costs in hand, we then pick
+ * the lower as the cost of the 'state<<1', and compute
+ * the path of bits leading to that state.
+ *
+ * Then, do the same for '(state<<1) + 1'. This time,
+ * instead of computing the m0 and m1 values from
+ * scratch, because the only difference is that we're
+ * expecting a one bit instead of a zero bit, we just
+ * flip the bitcost values around to match the
+ * expected transmitted bits with some tricky
+ * arithmetic which is equivalent to:
+ *
+ * m0 = cost[p][state] + (510 - bitcost);
+ * m1 = cost[p][state+4] + bitcost
+ *
+ * Then, the lowest cost and bit trace of the new state
+ * is saved.
*/
- for (state = 0; state < NUM_STATE; state++) {
- uint32_t bitcost = ((uint32_t) (s0 ^ ao_fec_decode_table[(state<<1)]) +
- (uint32_t) (s1 ^ ao_fec_decode_table[(state<<1)+1]));
- {
- uint32_t cost0 = cost[p][state] + bitcost;
- uint8_t state0 = ao_next_state(state, 0);
-
- if (cost0 < cost[n][state0]) {
- cost[n][state0] = cost0;
- bits[n][state0] = (bits[p][state] << 1) | (state & 1);
- }
- }
- {
- uint32_t cost1 = cost[p][state] + 510 - bitcost;
- uint8_t state1 = ao_next_state(state, 1);
- if (cost1 < cost[n][state1]) {
- cost[n][state1] = cost1;
- bits[n][state1] = (bits[p][state] << 1) | (state & 1);
- }
- }
+#define DO_STATE(state) { \
+ uint32_t bitcost; \
+ \
+ uint32_t m0; \
+ uint32_t m1; \
+ uint32_t bit; \
+ \
+ bitcost = ((uint32_t) (s0 ^ ao_fec_decode_table[(state<<1)]) + \
+ (uint32_t) (s1 ^ ao_fec_decode_table[(state<<1)|1])); \
+ \
+ m0 = cost[p][state] + bitcost; \
+ m1 = cost[p][state+4] + (510 - bitcost); \
+ bit = m0 > m1; \
+ cost[n][state<<1] = bit ? m1 : m0; \
+ bits[n][state<<1] = (bits[p][state + (bit<<2)] << 1) | (state&1); \
+ \
+ m0 -= (bitcost+bitcost-510); \
+ m1 += (bitcost+bitcost-510); \
+ bit = m0 > m1; \
+ cost[n][(state<<1)+1] = bit ? m1 : m0; \
+ bits[n][(state<<1)+1] = (bits[p][state + (bit<<2)] << 1) | (state&1); \
}
+ DO_STATE(0);
+ DO_STATE(1);
+ DO_STATE(2);
+ DO_STATE(3);
+
#if 0
printf ("bit %3d symbol %2x %2x:", i/2, s0, s1);
for (state = 0; state < NUM_STATE; state++) {
- printf (" %5d(%04x)", cost[n][state], bits[n][state]);
+ printf (" %8u(%08x)", cost[n][state], bits[n][state]);
}
printf ("\n");
#endif
int8_t dist = b - (o + 8); /* distance to last ready-for-writing bit */
uint32_t min_cost; /* lowest cost */
uint8_t min_state; /* lowest cost state */
+ uint8_t byte;
/* Find the best fit at the current point
* of the decode.
printf ("\tbit %3d min_cost %5d old bit %3d old_state %x bits %02x whiten %0x\n",
i/2, min_cost, o + 8, min_state, (bits[p][min_state] >> dist) & 0xff, *whiten);
#endif
- if (out_len) {
- uint8_t byte = (bits[p][min_state] >> dist) ^ *whiten++;
-
- if (out_len > 2) {
- crc = ao_fec_crc_byte(byte, crc);
- *out++ = byte;
- } else {
- *out++ = byte ^ (crc >> 8);
- crc <<= 8;
- }
- --out_len;
+ byte = (bits[p][min_state] >> dist) ^ *whiten++;
+ *out++ = byte;
+ if (out_len > 2)
+ crc = ao_fec_crc_byte(byte, crc);
+
+ if (!--out_len) {
+ if ((out[-2] == (uint8_t) (crc >> 8)) &&
+ out[-1] == (uint8_t) crc)
+ out[-1] = AO_FEC_DECODE_CRC_OK;
+ else
+ out[-1] = 0;
+ out[-2] = 0;
+ goto done;
}
o += 8;
}
}
+done:
#if AO_PROFILE
ao_fec_decode_start = start_tick;
ao_fec_decode_end = ao_profile_tick();
#endif
- return len/16;
+ return 1;
}
projects / fw / altos / blobdiff