[xiph-commits] r11700 - trunk/speex/libspeex
jm at svn.xiph.org
jm at svn.xiph.org
Mon Jul 10 04:44:08 PDT 2006
Author: jm
Date: 2006-07-10 04:44:05 -0700 (Mon, 10 Jul 2006)
New Revision: 11700
Modified:
trunk/speex/libspeex/ltp_bfin.h
Log:
Blackfin assembly patch by David Rowe
Modified: trunk/speex/libspeex/ltp_bfin.h
===================================================================
--- trunk/speex/libspeex/ltp_bfin.h 2006-07-09 06:20:12 UTC (rev 11699)
+++ trunk/speex/libspeex/ltp_bfin.h 2006-07-10 11:44:05 UTC (rev 11700)
@@ -152,3 +152,268 @@
return sum;
}
+#define OVERRIDE_OPEN_LOOP_NBEST_PITCH
+#ifdef OVERRIDE_OPEN_LOOP_NBEST_PITCH
+void open_loop_nbest_pitch(spx_word16_t *sw, int start, int end, int len, int *pitch, spx_word16_t *gain, int N, char *stack)
+{
+ int i,j,k;
+ VARDECL(spx_word32_t *best_score);
+ VARDECL(spx_word32_t *best_ener);
+ spx_word32_t e0;
+ VARDECL(spx_word32_t *corr);
+ VARDECL(spx_word32_t *energy);
+
+ ALLOC(best_score, N, spx_word32_t);
+ ALLOC(best_ener, N, spx_word32_t);
+ ALLOC(corr, end-start+1, spx_word32_t);
+ ALLOC(energy, end-start+2, spx_word32_t);
+
+ for (i=0;i<N;i++)
+ {
+ best_score[i]=-1;
+ best_ener[i]=0;
+ pitch[i]=start;
+ }
+
+ energy[0]=inner_prod(sw-start, sw-start, len);
+ e0=inner_prod(sw, sw, len);
+
+ /* energy update -------------------------------------*/
+
+ __asm__ __volatile__
+ (
+" P0 = %0;\n\t"
+" I1 = %1;\n\t"
+" L1 = 0;\n\t"
+" I2 = %2;\n\t"
+" L2 = 0;\n\t"
+" R2 = [P0++];\n\t"
+" R3 = 0;\n\t"
+" LSETUP (eu1, eu2) LC1 = %3;\n\t"
+"eu1: R1.L = W [I1--] || R0.L = W [I2--] ;\n\t"
+" R1 = R1.L * R1.L (IS);\n\t"
+" R0 = R0.L * R0.L (IS);\n\t"
+" R1 >>>= 6;\n\t"
+" R1 = R1 + R2;\n\t"
+" R0 >>>= 6;\n\t"
+" R1 = R1 - R0;\n\t"
+" R2 = MAX(R1,R3);\n\t"
+"eu2: [P0++] = R2;\n\t"
+ : : "d" (energy), "d" (&sw[-start-1]), "d" (&sw[-start+len-1]),
+ "a" (end-start)
+ : "P0", "I1", "I2", "R0", "R1", "R2", "R3"
+#if (__GNUC__ == 4)
+ , "LC1"
+#endif
+ );
+
+ pitch_xcorr(sw, sw-end, corr, len, end-start+1, stack);
+
+ /* FIXME: Fixed-point and floating-point code should be merged */
+ {
+ VARDECL(spx_word16_t *corr16);
+ VARDECL(spx_word16_t *ener16);
+ ALLOC(corr16, end-start+1, spx_word16_t);
+ ALLOC(ener16, end-start+1, spx_word16_t);
+ /* Normalize to 180 so we can square it and it still fits in 16 bits */
+ normalize16(corr, corr16, 180, end-start+1);
+ normalize16(energy, ener16, 180, end-start+1);
+
+ if (N == 1) {
+ /* optimised asm to handle N==1 case */
+ __asm__ __volatile__
+ (
+" I0 = %1;\n\t" /* I0: corr16[] */
+" L0 = 0;\n\t"
+" I1 = %2;\n\t" /* I1: energy */
+" L1 = 0;\n\t"
+" R2 = -1;\n\t" /* R2: best score */
+" R3 = 0;\n\t" /* R3: best energy */
+" P0 = %4;\n\t" /* P0: best pitch */
+" P1 = %4;\n\t" /* P1: counter */
+" LSETUP (sl1, sl2) LC1 = %3;\n\t"
+"sl1: R0.L = W [I0++] || R1.L = W [I1++];\n\t"
+" R0 = R0.L * R0.L (IS);\n\t"
+" R1 += 1;\n\t"
+" R4 = R0.L * R3.L;\n\t"
+" R5 = R2.L * R1.L;\n\t"
+" cc = R5 < R4;\n\t"
+" if cc R2 = R0;\n\t"
+" if cc R3 = R1;\n\t"
+" if cc P0 = P1;\n\t"
+"sl2: P1 += 1;\n\t"
+" %0 = P0;\n\t"
+ : "=&d" (pitch[0])
+ : "a" (corr16), "a" (ener16), "a" (end+1-start), "d" (start)
+ : "P0", "P1", "I0", "I1", "R0", "R1", "R2", "R3", "R4", "R5"
+#if (__GNUC__ == 4)
+ , "LC1"
+#endif
+ );
+
+ }
+ else {
+ for (i=start;i<=end;i++)
+ {
+ spx_word16_t tmp = MULT16_16_16(corr16[i-start],corr16[i-start]);
+ /* Instead of dividing the tmp by the energy, we multiply on the other side */
+ if (MULT16_16(tmp,best_ener[N-1])>MULT16_16(best_score[N-1],ADD16(1,ener16[i-start])))
+ {
+ /* We can safely put it last and then check */
+ best_score[N-1]=tmp;
+ best_ener[N-1]=ener16[i-start]+1;
+ pitch[N-1]=i;
+ /* Check if it comes in front of others */
+ for (j=0;j<N-1;j++)
+ {
+ if (MULT16_16(tmp,best_ener[j])>MULT16_16(best_score[j],ADD16(1,ener16[i-start])))
+ {
+ for (k=N-1;k>j;k--)
+ {
+ best_score[k]=best_score[k-1];
+ best_ener[k]=best_ener[k-1];
+ pitch[k]=pitch[k-1];
+ }
+ best_score[j]=tmp;
+ best_ener[j]=ener16[i-start]+1;
+ pitch[j]=i;
+ break;
+ }
+ }
+ }
+ }
+ }
+ }
+
+ /* Compute open-loop gain */
+ if (gain)
+ {
+ for (j=0;j<N;j++)
+ {
+ spx_word16_t g;
+ i=pitch[j];
+ g = DIV32(corr[i-start], 10+SHR32(MULT16_16(spx_sqrt(e0),spx_sqrt(energy[i-start])),6));
+ /* FIXME: g = max(g,corr/energy) */
+ if (g<0)
+ g = 0;
+ gain[j]=g;
+ }
+ }
+}
+#endif
+
+#define OVERRIDE_PITCH_GAIN_SEARCH_3TAP_VQ
+#ifdef OVERRIDE_PITCH_GAIN_SEARCH_3TAP_VQ
+static int pitch_gain_search_3tap_vq(
+ const signed char *gain_cdbk,
+ int gain_cdbk_size,
+ spx_word16_t *C16,
+ spx_word16_t max_gain
+)
+{
+ const signed char *ptr=gain_cdbk;
+ int best_cdbk=0;
+ spx_word32_t best_sum=-VERY_LARGE32;
+ spx_word32_t sum=0;
+ spx_word16_t g[3];
+ spx_word16_t pitch_control=64;
+ spx_word16_t gain_sum;
+ int i;
+
+ /* fast asm version of VQ codebook search */
+
+ __asm__ __volatile__
+ (
+
+" P0 = %2;\n\t" /* P0: ptr to gain_cdbk */
+" L1 = 0;\n\t" /* no circ addr for L1 */
+" %0 = 0;\n\t" /* %0: best_sum */
+" %1 = 0;\n\t" /* %1: best_cbdk */
+" P1 = 0;\n\t" /* P1: loop counter */
+" R5 = 64;\n\t" /* R5: pitch_control */
+
+" LSETUP (pgs1, pgs2) LC1 = %4;\n\t"
+"pgs1: R2 = B [P0++] (X);\n\t" /* R2: g[0] */
+" R3 = B [P0++] (X);\n\t" /* R3: g[1] */
+" R4 = B [P0++] (X);\n\t" /* R4: g[2] */
+" R2 += 32;\n\t"
+" R3 += 32;\n\t"
+" R4 += 32;\n\t"
+
+" R0 = B [P0++] (X);\n\t"
+" B0 = R0;\n\t" /* BO: gain_sum */
+
+ /* compute_pitch_error() -------------------------------*/
+
+" I1 = %3;\n\t" /* I1: ptr to C */
+" A0 = 0;\n\t"
+
+" R0.L = W[I1++];\n\t"
+" R1.L = R2.L*R5.L (IS);\n\t"
+" A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R3.L*R5.L (IS);\n\t"
+" A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R4.L*R5.L (IS);\n\t"
+" A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R2.L*R3.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R4.L*R3.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R4.L*R2.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R2.L*R2.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R3.L*R3.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R4.L*R4.L (IS);\n\t"
+" R0 = (A0 -= R1.L*R0.L) (IS);\n\t"
+
+/*
+ Re-arrange the if-then to code efficiently on the Blackfin:
+
+ if (sum>best_sum && gain_sum<=max_gain) ------ (1)
+
+ if (sum>best_sum && !(gain_sum>max_gain)) ------ (2)
+
+ if (max_gain<=gain_sum) { ------ (3)
+ sum = -VERY_LARGE32;
+ }
+ if (best_sum<=sum)
+
+ The blackin cc instructions are all of the form:
+
+ cc = x < y (or cc = x <= y)
+*/
+" R1 = B0\n\t"
+" R2 = %5\n\t"
+" R3 = %6\n\t"
+" cc = R2 <= R1;\n\t"
+" if cc R0 = R3;\n\t"
+" cc = %0 <= R0;\n\t"
+" if cc %0 = R0;\n\t"
+" if cc %1 = P1;\n\t"
+
+"pgs2: P1 += 1;\n\t"
+
+ : "=&d" (best_sum), "=&d" (best_cdbk)
+ : "a" (gain_cdbk), "a" (C16), "a" (gain_cdbk_size), "a" (max_gain),
+ "b" (-VERY_LARGE32)
+ : "R0", "R1", "R2", "R3", "R4", "R5", "P0",
+ "P1", "I1", "L1", "A0", "B0"
+#if (__GNUC__ == 4)
+ , "LC1"
+#endif
+ );
+
+ return best_cdbk;
+}
+#endif
+
More information about the commits
mailing list