QMF filter memory doesn't need to be 32 bits, so convert it to 16 bits. Alter assembler code to reflect this. Also do a small optimization in the reverse block copy of the qmf_synth() Coldfire assembler.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@15315 a1c6a512-1295-4272-9138-f99709370657

4 files changed, 21 insertions, 23 deletions

diff --git a/apps/codecs/libspeex/filters.c b/apps/codecs/libspeex/filters.c
index e64f087a5d..0e76e27e84 100644
--- a/apps/codecs/libspeex/filters.c
+++ b/apps/codecs/libspeex/filters.c
@@ -478,7 +478,7 @@ void qmf_decomp(const spx_word16_t *xx, const spx_word16_t *aa, spx_word16_t *y1
 
 #ifndef OVERRIDE_QMF_SYNTH
 /* Re-synthesised a signal from the QMF low-band and high-band signals */
-void qmf_synth(const spx_word16_t *x1, const spx_word16_t *x2, const spx_word16_t *a, spx_word16_t *y, int N, int M, spx_word32_t *mem1, spx_word32_t *mem2, char *stack)
+void qmf_synth(const spx_word16_t *x1, const spx_word16_t *x2, const spx_word16_t *a, spx_word16_t *y, int N, int M, spx_word16_t *mem1, spx_word16_t *mem2, char *stack)
    /* assumptions:
       all odd x[i] are zero -- well, actually they are left out of the array now
       N and M are multiples of 4 */
diff --git a/apps/codecs/libspeex/filters.h b/apps/codecs/libspeex/filters.h
index b363a9a683..b08e94d320 100644
--- a/apps/codecs/libspeex/filters.h
+++ b/apps/codecs/libspeex/filters.h
@@ -59,7 +59,7 @@ void highpass(const spx_word16_t *x, spx_word16_t *y, int len, int filtID, spx_m
 
 
 void qmf_decomp(const spx_word16_t *xx, const spx_word16_t *aa, spx_word16_t *, spx_word16_t *y2, int N, int M, spx_word16_t *mem, char *stack);
-void qmf_synth(const spx_word16_t *x1, const spx_word16_t *x2, const spx_word16_t *a, spx_word16_t *y, int N, int M, spx_word32_t *mem1, spx_word32_t *mem2, char *stack);
+void qmf_synth(const spx_word16_t *x1, const spx_word16_t *x2, const spx_word16_t *a, spx_word16_t *y, int N, int M, spx_word16_t *mem1, spx_word16_t *mem2, char *stack);
 
 void filter_mem16(const spx_word16_t *x, const spx_coef_t *num, const spx_coef_t *den, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack);
 void iir_mem16(const spx_word16_t *x, const spx_coef_t *den, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack);
diff --git a/apps/codecs/libspeex/filters_cf.S b/apps/codecs/libspeex/filters_cf.S
index c9aa742b6f..b0367025e1 100644
--- a/apps/codecs/libspeex/filters_cf.S
+++ b/apps/codecs/libspeex/filters_cf.S
@@ -194,30 +194,32 @@ qmf_synth:
     move.l   %a6, %sp                   | Update sp
     move.l   %d3, -(%sp)                | Stack old %sp
 
-    | Backwards copy x1 and x2 arrays to xx1 and xx2
+    | Backwards copy x1 and x2 arrays to xx1 and xx2, assume N2 is power of two
     | TODO: these copying loops probably have more potential for optimization
     lea.l    (%a0, %d0.l*2), %a0        | x1 += N2
     lea.l    (%a1, %d0.l*2), %a1        | x2 += N2
     move.l   %d0, %d2                   | Loop counter is N2
 0:
-    move.w   -(%a0), (%a2)+
-    move.w   -(%a1), (%a6)+
-    subq.l   #1, %d2
+    move.l   -(%a0), %d3
+    swap.w   %d3
+    move.l   %d3, (%a2)+
+    move.l   -(%a1), %d3
+    swap.w   %d3
+    move.l   %d3, (%a6)+
+    subq.l   #2, %d2
     jne      0b
 
     | Copy alternate members of mem1 and mem2 to last part of xx1 and xx2
     move.l   %d1, %d2                   | Loop counter is M2
-    addq.l   #4, %a4                    | a4 = &mem1[1]
-    addq.l   #4, %a5                    | a5 = &mem2[1]
+    addq.l   #2, %a4                    | a4 = &mem1[1]
+    addq.l   #2, %a5                    | a5 = &mem2[1]
     move.l   %a4, %d3                   | Backup mem1 and mem2
     move.l   %a5, %d4
 0:
-    move.l   (%a4), %d5
-    move.w   %d5, (%a2)+
-    move.l   (%a5), %d5
-    move.w   %d5, (%a6)+
-    addq.l   #8, %a4
-    addq.l   #8, %a5
+    move.w   (%a4), (%a2)+
+    move.w   (%a5), (%a6)+ 
+    addq.l   #4, %a4
+    addq.l   #4, %a5
     subq.l   #1, %d2
     jne      0b
     move.l   %d3, %a4                   | a4 = &mem1[1]
@@ -295,14 +297,10 @@ qmf_synth:
     addq.l   #4, %a0            | a0 = &xx1[0]
     addq.l   #4, %a1            | a1 = &xx2[0]
 0:
-    move.w   (%a0)+, %d2
-    move.w   (%a1)+, %d3
-    ext.l    %d2
-    ext.l    %d3
-    move.l   %d2, (%a4)
-    move.l   %d3, (%a5)
-    addq.l   #8, %a4
-    addq.l   #8, %a5
+    move.w   (%a0)+, (%a4)
+    move.w   (%a1)+, (%a5)
+    addq.l   #4, %a4
+    addq.l   #4, %a5
     subq.l   #1, %d1
     jne      0b
 
diff --git a/apps/codecs/libspeex/sb_celp.h b/apps/codecs/libspeex/sb_celp.h
index 9fb2241a83..da3545d943 100644
--- a/apps/codecs/libspeex/sb_celp.h
+++ b/apps/codecs/libspeex/sb_celp.h
@@ -109,7 +109,7 @@ typedef struct SBDecState {
    int    lpc_enh_enabled;
 
    char  *stack;
-   spx_word32_t g0_mem[64], g1_mem[64];
+   spx_word16_t g0_mem[64], g1_mem[64];
 
    spx_word16_t excBuf[80];
    spx_lsp_t old_qlsp[10];