Use more IRAM for mpc on S5L870x. Speed up decoding by ~2%.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@29891 a1c6a512-1295-4272-9138-f99709370657

5 files changed, 29 insertions, 5 deletions

diff --git a/apps/codecs/libmusepack/internal.h b/apps/codecs/libmusepack/internal.h
index 760f50b02d..0f29060a7d 100644
--- a/apps/codecs/libmusepack/internal.h
+++ b/apps/codecs/libmusepack/internal.h
@@ -58,7 +58,7 @@ typedef struct mpc_block_t {
 } mpc_block;
 
 #define MAX_FRAME_SIZE 4352
-#define DEMUX_BUFFER_SIZE (65536 - MAX_FRAME_SIZE) // need some space as sand box
+#define DEMUX_BUFFER_SIZE (32768 - MAX_FRAME_SIZE) // need some space as sand box
 
 struct mpc_demux_t {
     mpc_reader * r;
diff --git a/apps/codecs/libmusepack/mpc_decoder.c b/apps/codecs/libmusepack/mpc_decoder.c
index face3d1f81..eb0638524f 100644
--- a/apps/codecs/libmusepack/mpc_decoder.c
+++ b/apps/codecs/libmusepack/mpc_decoder.c
@@ -123,7 +123,8 @@ void mpc_decoder_requantisierung   (mpc_decoder *d)
 void mpc_decoder_read_bitstream_sv7(mpc_decoder * d, mpc_bits_reader * r) 
                                     ICODE_ATTR_MPC_LARGE_IRAM;
 void mpc_decoder_read_bitstream_sv8(mpc_decoder * d, mpc_bits_reader * r,
-                                    mpc_bool_t is_key_frame);
+                                    mpc_bool_t is_key_frame)
+                                    ICODE_ATTR_MPC_SV8_BS_DEC;
 
 //------------------------------------------------------------------------------
 // macros
diff --git a/apps/codecs/libmusepack/mpc_demux.c b/apps/codecs/libmusepack/mpc_demux.c
index e4ae0f89d3..499170c135 100644
--- a/apps/codecs/libmusepack/mpc_demux.c
+++ b/apps/codecs/libmusepack/mpc_demux.c
@@ -51,7 +51,7 @@
 #define MAX_BUFFER_SIZE (DEMUX_BUFFER_SIZE + MAX_FRAME_SIZE)
 
 // globals
-static mpc_uint8_t g_buffer[MAX_BUFFER_SIZE];
+static mpc_uint8_t g_buffer[MAX_BUFFER_SIZE] IBSS_ATTR_MPC_BITBUFFER;
 static mpc_seek_t g_seek_table[MAX_SEEK_TABLE_SIZE];
 static mpc_demux g_mpc_demux IBSS_ATTR;
 
diff --git a/apps/codecs/libmusepack/mpcdec.h b/apps/codecs/libmusepack/mpcdec.h
index d4c3bd22fc..68754dcbd2 100644
--- a/apps/codecs/libmusepack/mpcdec.h
+++ b/apps/codecs/libmusepack/mpcdec.h
@@ -52,23 +52,42 @@ extern "C" {
 #define IBSS_ATTR_MPC_LARGE_IRAM   IBSS_ATTR
 #define ICODE_ATTR_MPC_LARGE_IRAM
 #define ICONST_ATTR_MPC_LARGE_IRAM ICONST_ATTR
+/* Does not fit into IRAM. */
+#define IBSS_ATTR_MPC_BITBUFFER
+#define ICODE_ATTR_MPC_SV8_BS_DEC
 /* Keep the data arrays of bitsreadr.c in IRAM. */
 #define ICONST_ATTR_MPC_BITSREADER ICONST_ATTR
 
-#elif (CONFIG_CPU == PP5022) || (CONFIG_CPU == PP5024) || defined(CPU_S5L870X)
+#elif (CONFIG_CPU == PP5022) || (CONFIG_CPU == PP5024)
 /* Enough IRAM to move additional data and code to it. */
 #define IBSS_ATTR_MPC_LARGE_IRAM   IBSS_ATTR
 #define ICODE_ATTR_MPC_LARGE_IRAM  ICODE_ATTR
 #define ICONST_ATTR_MPC_LARGE_IRAM ICONST_ATTR
+/* Does not fit into IRAM. */
+#define IBSS_ATTR_MPC_BITBUFFER
+#define ICODE_ATTR_MPC_SV8_BS_DEC
 /* Not putting the data arrays of bitsreader.c to IRAM allows to move the
  * sv7/sv8 bitstream demuxing into IRAM. This config is faster. */
 #define ICONST_ATTR_MPC_BITSREADER
 
+#elif defined(CPU_S5L870X)
+/* Enough IRAM to move additional data and code to it. */
+#define IBSS_ATTR_MPC_LARGE_IRAM   IBSS_ATTR
+#define ICODE_ATTR_MPC_LARGE_IRAM  ICODE_ATTR
+#define ICONST_ATTR_MPC_LARGE_IRAM ICONST_ATTR
+/* Faster when moved to IRAM. */
+#define IBSS_ATTR_MPC_BITBUFFER    IBSS_ATTR
+#define ICODE_ATTR_MPC_SV8_BS_DEC  ICODE_ATTR
+/* Not faster when moved to IRAM. */
+#define ICONST_ATTR_MPC_BITSREADER
+
 #else
 /* Not enough IRAM available. */
 #define IBSS_ATTR_MPC_LARGE_IRAM
 #define ICODE_ATTR_MPC_LARGE_IRAM
 #define ICONST_ATTR_MPC_LARGE_IRAM
+#define IBSS_ATTR_MPC_BITBUFFER
+#define ICODE_ATTR_MPC_SV8_BS_DEC
 #define ICONST_ATTR_MPC_BITSREADER
 #endif
 
diff --git a/apps/codecs/libmusepack/synth_filter.c b/apps/codecs/libmusepack/synth_filter.c
index 602edff6c6..3ad2b236a0 100644
--- a/apps/codecs/libmusepack/synth_filter.c
+++ b/apps/codecs/libmusepack/synth_filter.c
@@ -510,7 +510,11 @@ mpc_decoder_windowing_D(MPC_SAMPLE_FORMAT * Data,
 #endif /* CPU_ARM || CPU_COLDFIRE */
 
 static void 
-mpc_full_synthesis_filter(MPC_SAMPLE_FORMAT *OutData, MPC_SAMPLE_FORMAT *V, const MPC_SAMPLE_FORMAT *Y)
+mpc_full_synthesis_filter(MPC_SAMPLE_FORMAT *OutData, MPC_SAMPLE_FORMAT *V, 
+                          const MPC_SAMPLE_FORMAT *Y) ICODE_ATTR_MPC_LARGE_IRAM;
+static void 
+mpc_full_synthesis_filter(MPC_SAMPLE_FORMAT *OutData, MPC_SAMPLE_FORMAT *V, 
+                          const MPC_SAMPLE_FORMAT *Y)
 {
     mpc_uint32_t n;