Unroll mpc's requantization loops for some (minor) speed up.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@25891 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 60 insertions, 19 deletions

diff --git a/apps/codecs/libmusepack/mpc_decoder.c b/apps/codecs/libmusepack/mpc_decoder.c
index 9e4ef634bf..f403673818 100644
--- a/apps/codecs/libmusepack/mpc_decoder.c
+++ b/apps/codecs/libmusepack/mpc_decoder.c
@@ -125,54 +125,95 @@ void mpc_decoder_read_bitstream_sv8(mpc_decoder * d, mpc_bits_reader * r,
 //------------------------------------------------------------------------------
 // macros
 //------------------------------------------------------------------------------
+#define REQUANT_M1_S1_SAMPLES(IDX) \
+    *(YL+=IDX) = (templ = MPC_MULTIPLY_FLOAT_INT(facL,*L++))+(tempr = MPC_MULTIPLY_FLOAT_INT(facR,*R++)); \
+    *(YR+=IDX) = templ - tempr;
+
 #define REQUANT_M1_S1(SUBFRAME) \
     facL = MPC_MULTIPLY_SCF( Cc[d->Res_L[Band]] , d->SCF_Index_L[Band][SUBFRAME] & 0xFF); \
     facR = MPC_MULTIPLY_SCF( Cc[d->Res_R[Band]] , d->SCF_Index_R[Band][SUBFRAME] & 0xFF); \
-    for (n = 0; n < 12; n++, YL += 32, YR += 32) { \
-        *YL   = (templ = MPC_MULTIPLY_FLOAT_INT(facL,*L++))+(tempr = MPC_MULTIPLY_FLOAT_INT(facR,*R++)); \
-        *YR   = templ - tempr; \
+    for (n = 0; n < 12; n+=4, YL += 32, YR += 32) { \
+        REQUANT_M1_S1_SAMPLES( 0); \
+        REQUANT_M1_S1_SAMPLES(32); \
+        REQUANT_M1_S1_SAMPLES(32); \
+        REQUANT_M1_S1_SAMPLES(32); \
     }
 
+#define REQUANT_M1_S0_SAMPLES(IDX) \
+    *(YR+=IDX) = *(YL+=IDX) = MPC_MULTIPLY_FLOAT_INT(facL,*L++);
+
 #define REQUANT_M1_S0(SUBFRAME) \
     facL = MPC_MULTIPLY_SCF( Cc[d->Res_L[Band]] , d->SCF_Index_L[Band][SUBFRAME] & 0xFF); \
-    for (n = 0; n < 12; n++, YL += 32, YR += 32) { \
-        *YR = *YL = MPC_MULTIPLY_FLOAT_INT(facL,*L++); \
+    for (n = 0; n < 12; n+=4, YL += 32, YR += 32) { \
+        REQUANT_M1_S0_SAMPLES( 0); \
+        REQUANT_M1_S0_SAMPLES(32); \
+        REQUANT_M1_S0_SAMPLES(32); \
+        REQUANT_M1_S0_SAMPLES(32); \
     }
 
+#define REQUANT_M0_S1_SAMPLES(IDX) \
+    *(YR+=IDX) = -(*(YL+=IDX) = MPC_MULTIPLY_FLOAT_INT(facR,*R++));
+
 #define REQUANT_M0_S1(SUBFRAME) \
     facR = MPC_MULTIPLY_SCF( Cc[d->Res_R[Band]] , d->SCF_Index_R[Band][SUBFRAME] & 0xFF); \
-    for (n = 0; n < 12; n++, YL += 32, YR += 32) { \
-        *YR = - (*YL = MPC_MULTIPLY_FLOAT_INT(facR,*R++)); \
+    for (n = 0; n < 12; n+=4, YL += 32, YR += 32) { \
+        REQUANT_M0_S1_SAMPLES( 0); \
+        REQUANT_M0_S1_SAMPLES(32); \
+        REQUANT_M0_S1_SAMPLES(32); \
+        REQUANT_M0_S1_SAMPLES(32); \
     }
 
+#define REQUANT_L1_R1_SAMPLES(IDX) \
+    *(YL+=IDX) = MPC_MULTIPLY_FLOAT_INT(facL,*L++); \
+    *(YR+=IDX) = MPC_MULTIPLY_FLOAT_INT(facR,*R++);
+
 #define REQUANT_L1_R1(SUBFRAME) \
     facL = MPC_MULTIPLY_SCF( Cc[d->Res_L[Band]] , d->SCF_Index_L[Band][SUBFRAME] & 0xFF); \
     facR = MPC_MULTIPLY_SCF( Cc[d->Res_R[Band]] , d->SCF_Index_R[Band][SUBFRAME] & 0xFF); \
-    for (n = 0; n < 12; n++, YL += 32, YR += 32) { \
-        *YL = MPC_MULTIPLY_FLOAT_INT(facL,*L++); \
-        *YR = MPC_MULTIPLY_FLOAT_INT(facR,*R++); \
+    for (n = 0; n < 12; n+=4, YL += 32, YR += 32) { \
+        REQUANT_L1_R1_SAMPLES( 0); \
+        REQUANT_L1_R1_SAMPLES(32); \
+        REQUANT_L1_R1_SAMPLES(32); \
+        REQUANT_L1_R1_SAMPLES(32); \
     }
 
+#define REQUANT_L1_R0_SAMPLES(IDX) \
+    *(YL+=IDX) = MPC_MULTIPLY_FLOAT_INT(facL,*L++); \
+    *(YR+=IDX) = 0;
+
 #define REQUANT_L1_R0(SUBFRAME) \
     facL = MPC_MULTIPLY_SCF( Cc[d->Res_L[Band]] , d->SCF_Index_L[Band][SUBFRAME] & 0xFF); \
-    for (n = 0; n < 12; n++, YL += 32, YR += 32) { \
-        *YL = MPC_MULTIPLY_FLOAT_INT(facL,*L++); \
-        *YR = 0; \
+    for (n = 0; n < 12; n+=4, YL += 32, YR += 32) { \
+        REQUANT_L1_R0_SAMPLES( 0); \
+        REQUANT_L1_R0_SAMPLES(32); \
+        REQUANT_L1_R0_SAMPLES(32); \
+        REQUANT_L1_R0_SAMPLES(32); \
     }
 
+#define REQUANT_L0_R1_SAMPLES(IDX) \
+    *(YL+=IDX) = 0; \
+    *(YR+=IDX) = MPC_MULTIPLY_FLOAT_INT(facR,*R++);
+
 #define REQUANT_L0_R1(SUBFRAME) \
     facR = MPC_MULTIPLY_SCF( Cc[d->Res_R[Band]] , d->SCF_Index_R[Band][SUBFRAME] & 0xFF); \
-    for (n = 0; n < 12; n++, YL += 32, YR += 32) { \
-        *YL = 0; \
-        *YR = MPC_MULTIPLY_FLOAT_INT(facR,*R++); \
+    for (n = 0; n < 12; n+=4, YL += 32, YR += 32) { \
+        REQUANT_L0_R1_SAMPLES( 0); \
+        REQUANT_L0_R1_SAMPLES(32); \
+        REQUANT_L0_R1_SAMPLES(32); \
+        REQUANT_L0_R1_SAMPLES(32); \
     }
 
+#define REQUANT_SILENCE_SAMPLES(IDX) \
+    *(YR+=IDX) = *(YL+=IDX) = 0;
+
 #define REQUANT_SILENCE \
-    for (n = 0; n < 36; n++, YL += 32, YR += 32) { \
-        *YR = *YL = 0; \
+    for (n = 0; n < 12; n+=4, YL += 32, YR += 32) { \
+        REQUANT_SILENCE_SAMPLES( 0); \
+        REQUANT_SILENCE_SAMPLES(32); \
+        REQUANT_SILENCE_SAMPLES(32); \
+        REQUANT_SILENCE_SAMPLES(32); \
     }
 
-
 /**
  * set the scf indexes for seeking use
  * needed only for sv7 seeking