altos: Unroll viterbi state loop for >30% performance boost

[fw/altos] / src / core / ao_fec_rx.c

diff --git a/src/core/ao_fec_rx.c b/src/core/ao_fec_rx.c
index 0d400bb0b882ae90dd69c46eb58e9412e7347579..c1a3918b9ba53a3fdc952155cb62670027f84fd8 100644 (file)
--- a/src/core/ao_fec_rx.c
+++ b/src/core/ao_fec_rx.c
@@ -161,31 +161,43 @@ ao_fec_decode(const uint8_t *in, uint16_t len, uint8_t *out, uint8_t out_len, ui
                        cost[n][state] = 0x7fffffff;
 
                /* Compute path costs and accumulate output bit path
-                * for each state and encoded bit value
+                * 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
                 */
-               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) (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); \
+                               }                                       \
+                       }                                               \
                }
 
+               DO_STATE(0);
+               DO_STATE(1);
+               DO_STATE(2);
+               DO_STATE(3);
+               DO_STATE(4);
+               DO_STATE(5);
+               DO_STATE(6);
+               DO_STATE(7);
+
 #if 0
                printf ("bit %3d symbol %2x %2x:", i/2, s0, s1);
                for (state = 0; state < NUM_STATE; state++) {
@@ -211,6 +223,7 @@ ao_fec_decode(const uint8_t *in, uint16_t len, uint8_t *out, uint8_t out_len, ui
                        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.
@@ -238,24 +251,27 @@ ao_fec_decode(const uint8_t *in, uint16_t len, uint8_t *out, uint8_t out_len, ui
                        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++;
+                       byte = (bits[p][min_state] >> dist) ^ *whiten++;
+                       *out++ = byte;
+                       if (out_len > 2)
+                               crc = ao_fec_crc_byte(byte, crc);
 
-                               if (out_len > 2) {
-                                       crc = ao_fec_crc_byte(byte, crc);
-                                       *out++ = byte;
-                               } else {
-                                       *out++ = byte ^ (crc >> 8);
-                                       crc <<= 8;
-                               }
-                               --out_len;
+                       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;
 }