Speed up AAC-HE SBR by 2% on S5L8701. Use MEM_ALIGN on critical arrays and avoid stalls in asm code.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@29209 a1c6a512-1295-4272-9138-f99709370657

3 files changed, 33 insertions, 33 deletions

diff --git a/apps/codecs/libfaad/sbr_dct.c b/apps/codecs/libfaad/sbr_dct.c
index f22a24a18e..da0e5e1f59 100644
--- a/apps/codecs/libfaad/sbr_dct.c
+++ b/apps/codecs/libfaad/sbr_dct.c
@@ -1453,7 +1453,7 @@ void DCT2_32_unscaled(real_t *y, real_t *x)
 #else /* #ifdef SBR_LOW_POWER */
 
 /* table for pre-processing within dct4_kernel() */
-static const real_t dct4_pre_tab[] ICONST_ATTR = {
+static const real_t dct4_pre_tab[] ICONST_ATTR MEM_ALIGN_ATTR = {
     COEF_CONST(0.999924719333649), COEF_CONST(-1.01219630241394), COEF_CONST(-0.987653195858002),
     COEF_CONST(0.998118102550507), COEF_CONST(-1.05943882465363), COEF_CONST(-0.936797380447388),
     COEF_CONST(0.993906974792480), COEF_CONST(-1.10412919521332), COEF_CONST(-0.883684754371643),
@@ -1489,7 +1489,7 @@ static const real_t dct4_pre_tab[] ICONST_ATTR = {
 };
 
 /* table for post-processing within dct4_kernel() */
-static const real_t dct4_post_tab[] ICONST_ATTR = {
+static const real_t dct4_post_tab[] ICONST_ATTR MEM_ALIGN_ATTR = {
     COEF_CONST(1                ), COEF_CONST(-1               ), COEF_CONST(-1                ),
     COEF_CONST(0.998795449733734), COEF_CONST(-1.04786312580109), COEF_CONST(-0.949727773666382),
     COEF_CONST(0.995184719562531), COEF_CONST(-1.09320187568665), COEF_CONST(-0.897167563438416),
@@ -1525,7 +1525,7 @@ static const real_t dct4_post_tab[] ICONST_ATTR = {
 };
 
 // Table adapted from codeclib to fit into IRAM
-const uint32_t dct4_revtab[32] ICONST_ATTR = {
+const uint32_t dct4_revtab[32] ICONST_ATTR MEM_ALIGN_ATTR = {
   0, 24, 12, 22, 6, 30, 11, 19, 3, 27, 15, 21, 5, 29, 9, 17,
   1, 25, 13, 23, 7, 31, 10, 18, 2, 26, 14, 20, 4, 28, 8, 16};
 
diff --git a/apps/codecs/libfaad/sbr_dec.c b/apps/codecs/libfaad/sbr_dec.c
index 60bb2a6bdb..78c9c79d5d 100644
--- a/apps/codecs/libfaad/sbr_dec.c
+++ b/apps/codecs/libfaad/sbr_dec.c
@@ -521,8 +521,8 @@ uint8_t sbrDecodeSingleFrame(sbr_info *sbr, real_t *channel,
     return 0;
 }
 
-ALIGN qmf_t X_left[MAX_NTSRHFG][64];// = {{0}};
-ALIGN qmf_t X_right[MAX_NTSRHFG][64];// = {{0}}; /* must set this to 0 */
+qmf_t X_left [MAX_NTSRHFG][64] MEM_ALIGN_ATTR;// = {{0}};
+qmf_t X_right[MAX_NTSRHFG][64] MEM_ALIGN_ATTR;// = {{0}}; /* must set this to 0 */
 
 #if (defined(PS_DEC) || defined(DRM_PS))
 uint8_t sbrDecodeSingleFramePS(sbr_info *sbr, real_t *left_channel, real_t *right_channel,
diff --git a/apps/codecs/libfaad/sbr_qmf.c b/apps/codecs/libfaad/sbr_qmf.c
index f94c3072a4..bb6e176a3f 100644
--- a/apps/codecs/libfaad/sbr_qmf.c
+++ b/apps/codecs/libfaad/sbr_qmf.c
@@ -78,12 +78,12 @@ void qmfa_end(qmfa_info *qmfa)
 void sbr_qmf_analysis_32(sbr_info *sbr, qmfa_info *qmfa, const real_t *input,
                          qmf_t X[MAX_NTSRHFG][64], uint8_t offset, uint8_t kx)
 {
-    ALIGN real_t u[64];
+    real_t u[64] MEM_ALIGN_ATTR;
 #ifndef SBR_LOW_POWER
-    ALIGN real_t real[32];
-    ALIGN real_t imag[32];
+    real_t real[32] MEM_ALIGN_ATTR;
+    real_t imag[32] MEM_ALIGN_ATTR;
 #else
-    ALIGN real_t y[32];
+    real_t y[32] MEM_ALIGN_ATTR;
 #endif
     qmf_t *pX;
     uint32_t in = 0;
@@ -227,8 +227,8 @@ void qmfs_end(qmfs_info *qmfs)
 void sbr_qmf_synthesis_32(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
                           real_t *output)
 {
-    ALIGN real_t x[16];
-    ALIGN real_t y[16];
+    real_t x[16] MEM_ALIGN_ATTR;
+    real_t y[16] MEM_ALIGN_ATTR;
     int16_t n, k, out = 0;
     uint8_t l;
 
@@ -291,8 +291,8 @@ void sbr_qmf_synthesis_32(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][6
 void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
                           real_t *output)
 {
-    ALIGN real_t x[64];
-    ALIGN real_t y[64];
+    real_t x[64] MEM_ALIGN_ATTR;
+    real_t y[64] MEM_ALIGN_ATTR;
     int16_t n, k, out = 0;
     uint8_t l;
 
@@ -401,8 +401,8 @@ static const complex_t qmf32_pre_twiddle[] =
 void sbr_qmf_synthesis_32(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
                           real_t *output)
 {
-    ALIGN real_t x1[32];
-    ALIGN real_t x2[32];
+    real_t x1[32] MEM_ALIGN_ATTR;
+    real_t x2[32] MEM_ALIGN_ATTR;
     int32_t n, k, idx0, idx1, out = 0;
     uint32_t l;
 
@@ -464,10 +464,10 @@ void sbr_qmf_synthesis_32(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][6
 void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
                           real_t *output)
 {
-    ALIGN real_t real1[32]; 
-    ALIGN real_t imag1[32];
-    ALIGN real_t real2[32]; 
-    ALIGN real_t imag2[32];
+    real_t real1[32] MEM_ALIGN_ATTR; 
+    real_t imag1[32] MEM_ALIGN_ATTR;
+    real_t real2[32] MEM_ALIGN_ATTR; 
+    real_t imag2[32] MEM_ALIGN_ATTR;
     qmf_t *pX;
     real_t *p_buf_1, *p_buf_3;
     int32_t n, k, idx0, idx1, out = 0;
@@ -517,36 +517,36 @@ void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][6
             asm volatile (
                "ldmia %[qtab]!, { r0-r3 } \n\t"
                "ldr r4, [%[pbuf]]         \n\t"
+               "ldr r7, [%[pbuf], #192*4] \n\t"
                "smull r5, r6, r4, r0      \n\t"
-               "ldr r4, [%[pbuf], #192*4] \n\t"
-               "smlal r5, r6, r4, r1      \n\t"
                "ldr r4, [%[pbuf], #256*4] \n\t"
+               "smlal r5, r6, r7, r1      \n\t"
+               "ldr r7, [%[pbuf], #448*4] \n\t"
                "smlal r5, r6, r4, r2      \n\t"
-               "ldr r4, [%[pbuf], #448*4] \n\t"
-               "smlal r5, r6, r4, r3      \n\t"
+               "ldr r4, [%[pbuf], #512*4] \n\t"
+               "smlal r5, r6, r7, r3      \n\t"
 
                "ldmia %[qtab]!, { r0-r3 } \n\t"
-               "ldr r4, [%[pbuf], #512*4] \n\t"
+               "ldr r7, [%[pbuf], #704*4] \n\t"
                "smlal r5, r6, r4, r0      \n\t"
-               "ldr r4, [%[pbuf], #704*4] \n\t"
-               "smlal r5, r6, r4, r1      \n\t"
                "ldr r4, [%[pbuf], #768*4] \n\t"
+               "smlal r5, r6, r7, r1      \n\t"
+               "ldr r7, [%[pbuf], #960*4] \n\t"
                "smlal r5, r6, r4, r2      \n\t"
-               "ldr r4, [%[pbuf], #960*4] \n\t"
-               "smlal r5, r6, r4, r3      \n\t"
+               "mov r2, #1024*4           \n\t"
 
                "ldmia %[qtab]!, { r0-r1 } \n\t"
-               "mov r2, #1024*4           \n\t"
                "ldr r4, [%[pbuf], r2]     \n\t"
-               "smlal r5, r6, r4, r0      \n\t"
+               "smlal r5, r6, r7, r3      \n\t"
                "mov r2, #1216*4           \n\t"
-               "ldr r4, [%[pbuf], r2]     \n\t"
-               "smlal r5, r6, r4, r1      \n\t"
+               "ldr r7, [%[pbuf], r2]     \n\t"
+               "smlal r5, r6, r4, r0      \n\t"
+               "smlal r5, r6, r7, r1      \n\t"
 
                "str r6, [%[pout]]         \n"
                : [qtab] "+r" (qtab)
                : [pbuf] "r" (pbuf), [pout] "r" (pout)
-               : "r0", "r1", "r2", "r3", "r4", "r5", "r6", "memory");
+               : "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7", "memory");
         }
 #elif defined CPU_COLDFIRE
         const real_t *qtab = qmf_c;