Merge from branches/mdctexp - faster ifft+imdct in codec lib

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24712 a1c6a512-1295-4272-9138-f99709370657

4 files changed, 9 insertions, 170 deletions

diff --git a/apps/codecs/libwma/wmadec.h b/apps/codecs/libwma/wmadec.h
index a547ece157..4efaa9b8a2 100644
--- a/apps/codecs/libwma/wmadec.h
+++ b/apps/codecs/libwma/wmadec.h
@@ -23,8 +23,6 @@
 #include "asf.h"
 #include "bitstream.h" /* For GetBitContext */
 #include "types.h"
-//#include "dsputil.h"  /* For MDCTContext */
-
 
 //#define TRACE
 /* size of blocks */
diff --git a/apps/codecs/libwma/wmadeci.c b/apps/codecs/libwma/wmadeci.c
index ae1a93ecf2..6ff6a176ee 100644
--- a/apps/codecs/libwma/wmadeci.c
+++ b/apps/codecs/libwma/wmadeci.c
@@ -452,17 +452,6 @@ int wma_decode_init(WMADecodeContext* s, asf_waveformatex_t *wfx)
         }
     }
 
-    /*Not using the ffmpeg IMDCT anymore*/
-
-    /* mdct_init_global();
-
-    for(i = 0; i < s->nb_block_sizes; ++i)
-    {
-        ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
-
-    }
-    */
-
     /* ffmpeg uses malloc to only allocate as many window sizes as needed.  
     *  However, we're really only interested in the worst case memory usage.
     *  In the worst case you can have 5 window sizes, 128 doubling up 2048
@@ -1253,14 +1242,9 @@ static int wma_decode_block(WMADecodeContext *s, int32_t *scratch_buffer)
 
             n4 = s->block_len >>1;
 
-            /*faster IMDCT from Vorbis*/
-            mdct_backward( (1 << (s->block_len_bits+1)), (int32_t*)(*(s->coefs))[ch], (int32_t*)scratch_buffer);
-
-            /*slower but more easily understood IMDCT from FFMPEG*/
-            //ff_imdct_calc(&s->mdct_ctx[bsize],
-            //              output,
-            //              (*(s->coefs))[ch]);
-
+            ff_imdct_calc( (s->frame_len_bits - bsize + 1),
+                          (int32_t*)scratch_buffer,
+                          (*(s->coefs))[ch]);
 
             /* add in the frame */
             index = (s->frame_len / 2) + s->block_pos - n4;
diff --git a/apps/codecs/libwma/wmafixed.c b/apps/codecs/libwma/wmafixed.c
index 5569309145..1472ed081c 100644
--- a/apps/codecs/libwma/wmafixed.c
+++ b/apps/codecs/libwma/wmafixed.c
@@ -250,113 +250,3 @@ fixed64 fixdiv64(fixed64 x, fixed64 y)
 
     return (fixed32)(r << (PRECISION / 2));
 }
-
-
-
-/* Inverse gain of circular cordic rotation in s0.31 format. */
-static const long cordic_circular_gain = 0xb2458939; /* 0.607252929 */
-
-/* Table of values of atan(2^-i) in 0.32 format fractions of pi where pi = 0xffffffff / 2 */
-static const unsigned long atan_table[] = {
-    0x1fffffff, /* +0.785398163 (or pi/4) */
-    0x12e4051d, /* +0.463647609 */
-    0x09fb385b, /* +0.244978663 */
-    0x051111d4, /* +0.124354995 */
-    0x028b0d43, /* +0.062418810 */
-    0x0145d7e1, /* +0.031239833 */
-    0x00a2f61e, /* +0.015623729 */
-    0x00517c55, /* +0.007812341 */
-    0x0028be53, /* +0.003906230 */
-    0x00145f2e, /* +0.001953123 */
-    0x000a2f98, /* +0.000976562 */
-    0x000517cc, /* +0.000488281 */
-    0x00028be6, /* +0.000244141 */
-    0x000145f3, /* +0.000122070 */
-    0x0000a2f9, /* +0.000061035 */
-    0x0000517c, /* +0.000030518 */
-    0x000028be, /* +0.000015259 */
-    0x0000145f, /* +0.000007629 */
-    0x00000a2f, /* +0.000003815 */
-    0x00000517, /* +0.000001907 */
-    0x0000028b, /* +0.000000954 */
-    0x00000145, /* +0.000000477 */
-    0x000000a2, /* +0.000000238 */
-    0x00000051, /* +0.000000119 */
-    0x00000028, /* +0.000000060 */
-    0x00000014, /* +0.000000030 */
-    0x0000000a, /* +0.000000015 */
-    0x00000005, /* +0.000000007 */
-    0x00000002, /* +0.000000004 */
-    0x00000001, /* +0.000000002 */
-    0x00000000, /* +0.000000001 */
-    0x00000000, /* +0.000000000 */
-};
-
-
-/*
-
-    Below here functions do not use standard fixed precision!
-*/
-
-
-/**
- * Implements sin and cos using CORDIC rotation.
- *
- * @param phase has range from 0 to 0xffffffff, representing 0 and
- *        2*pi respectively.
- * @param cos return address for cos
- * @return sin of phase, value is a signed value from LONG_MIN to LONG_MAX,
- *         representing -1 and 1 respectively.
- *
- *        Gives at least 24 bits precision (last 2-8 bits or so are probably off)
- */
-long fsincos(unsigned long phase, fixed32 *cos)
-{
-    int32_t x, x1, y, y1;
-    unsigned long z, z1;
-    int i;
-
-    /* Setup initial vector */
-    x = cordic_circular_gain;
-    y = 0;
-    z = phase;
-
-    /* The phase has to be somewhere between 0..pi for this to work right */
-    if (z < 0xffffffff / 4) {
-        /* z in first quadrant, z += pi/2 to correct */
-        x = -x;
-        z += 0xffffffff / 4;
-    } else if (z < 3 * (0xffffffff / 4)) {
-        /* z in third quadrant, z -= pi/2 to correct */
-        z -= 0xffffffff / 4;
-    } else {
-        /* z in fourth quadrant, z -= 3pi/2 to correct */
-        x = -x;
-        z -= 3 * (0xffffffff / 4);
-    }
-
-    /* Each iteration adds roughly 1-bit of extra precision */
-    for (i = 0; i < 31; i++) {
-        x1 = x >> i;
-        y1 = y >> i;
-        z1 = atan_table[i];
-
-        /* Decided which direction to rotate vector. Pivot point is pi/2 */
-        if (z >= 0xffffffff / 4) {
-            x -= y1;
-            y += x1;
-            z -= z1;
-        } else {
-            x += y1;
-            y -= x1;
-            z += z1;
-        }
-    }
-
-    if (cos)
-        *cos = x;
-
-    return y;
-}
-
-
diff --git a/apps/codecs/libwma/wmafixed.h b/apps/codecs/libwma/wmafixed.h
index 6b5137e044..0ecdc5cfbc 100644
--- a/apps/codecs/libwma/wmafixed.h
+++ b/apps/codecs/libwma/wmafixed.h
@@ -52,10 +52,10 @@ fixed64 fixdiv64(fixed64 x, fixed64 y);
 fixed32 fixsqrt32(fixed32 x);
 long fsincos(unsigned long phase, fixed32 *cos);
 
+
 #ifdef CPU_ARM
 
 /*Sign-15.16 format */
-
 #define fixmul32(x, y)  \
     ({ int32_t __hi;  \
        uint32_t __lo;  \
@@ -70,18 +70,6 @@ long fsincos(unsigned long phase, fixed32 *cos);
        __result;  \
     })
 
-#define fixmul32b(x, y)  \
-    ({ int32_t __hi;  \
-       uint32_t __lo;  \
-       int32_t __result;  \
-       asm ("smull   %0, %1, %3, %4\n\t"  \
-            "movs    %2, %1, lsl #1"  \
-            : "=&r" (__lo), "=&r" (__hi), "=r" (__result)  \
-            : "%r" (x), "r" (y)  \
-            : "cc");  \
-       __result;  \
-    })
-
 #elif defined(CPU_COLDFIRE)
 
 static inline int32_t fixmul32(int32_t x, int32_t y)
@@ -91,9 +79,9 @@ static inline int32_t fixmul32(int32_t x, int32_t y)
 #endif
     int32_t t1;
     asm (
-        "mac.l   %[x], %[y], %%acc0  \n" /* multiply */
-        "mulu.l  %[y], %[x]      \n"     /* get lower half, avoid emac stall */
-        "movclr.l %%acc0, %[t1]  \n"     /* get higher half */
+        "mac.l   %[x], %[y], %%acc0  \n" // multiply
+        "mulu.l  %[y], %[x]      \n"     // get lower half, avoid emac stall
+        "movclr.l %%acc0, %[t1]  \n"     // get higher half
         "lsr.l   #1, %[t1]       \n"
         "move.w  %[t1], %[x]     \n"
         "swap    %[x]            \n"
@@ -103,17 +91,6 @@ static inline int32_t fixmul32(int32_t x, int32_t y)
     return x;
 }
 
-static inline int32_t fixmul32b(int32_t x, int32_t y)
-{
-    asm (
-        "mac.l   %[x], %[y], %%acc0  \n" /* multiply */
-        "movclr.l %%acc0, %[x]  \n"     /* get higher half */
-        : [x] "+d" (x)
-        : [y] "d"  (y)
-    );
-    return x;
-}
-
 #else
 
 static inline fixed32 fixmul32(fixed32 x, fixed32 y)
@@ -127,17 +104,7 @@ static inline fixed32 fixmul32(fixed32 x, fixed32 y)
     return (fixed32)temp;
 }
 
-static inline fixed32 fixmul32b(fixed32 x, fixed32 y)
-{
-    fixed64 temp;
-
-    temp = x;
-    temp *= y;
-
-    temp >>= 31;        //16+31-16 = 31 bits
-
-    return (fixed32)temp;
-}
-
 #endif
 
+
+/* get fixmul32b from codeclib */