dca: ARMv6 optimised decode_blockcode()

This is a hand-tuned version of the code with impossible parts of
the FASTDIV function ommitted.

2-5% faster overall on Cortex-A8.
Signed-off-by: Mans Rullgard <mans@mansr.com>

libavcodec/arm/dca.h

@@ -23,6 +23,61 @@
 
 #include <stdint.h>
 #include "config.h"
+#include "libavutil/intmath.h"
+
+#if HAVE_ARMV6 && HAVE_INLINE_ASM
+
+#define decode_blockcodes decode_blockcodes
+static inline int decode_blockcodes(int code1, int code2, int levels,
+                                    int *values)
+{
+    int v0, v1, v2, v3, v4, v5;
+
+    __asm__ ("smmul   %8,  %14, %18           \n"
+             "smmul   %11, %15, %18           \n"
+             "smlabb  %14, %8,  %17, %14      \n"
+             "smlabb  %15, %11, %17, %15      \n"
+             "smmul   %9,  %8,  %18           \n"
+             "smmul   %12, %11, %18           \n"
+             "sub     %14, %14, %16, lsr #1   \n"
+             "sub     %15, %15, %16, lsr #1   \n"
+             "smlabb  %8,  %9,  %17, %8       \n"
+             "smlabb  %11, %12, %17, %11      \n"
+             "smmul   %10, %9,  %18           \n"
+             "smmul   %13, %12, %18           \n"
+             "str     %14, %0                 \n"
+             "str     %15, %4                 \n"
+             "sub     %8,  %8,  %16, lsr #1   \n"
+             "sub     %11, %11, %16, lsr #1   \n"
+             "smlabb  %9,  %10, %17, %9       \n"
+             "smlabb  %12, %13, %17, %12      \n"
+             "smmul   %14, %10, %18           \n"
+             "smmul   %15, %13, %18           \n"
+             "str     %8,  %1                 \n"
+             "str     %11, %5                 \n"
+             "sub     %9,  %9,  %16, lsr #1   \n"
+             "sub     %12, %12, %16, lsr #1   \n"
+             "smlabb  %10, %14, %17, %10      \n"
+             "smlabb  %13, %15, %17, %13      \n"
+             "str     %9,  %2                 \n"
+             "str     %12, %6                 \n"
+             "sub     %10, %10, %16, lsr #1   \n"
+             "sub     %13, %13, %16, lsr #1   \n"
+             "str     %10, %3                 \n"
+             "str     %13, %7                 \n"
+             : "=m"(values[0]), "=m"(values[1]),
+               "=m"(values[2]), "=m"(values[3]),
+               "=m"(values[4]), "=m"(values[5]),
+               "=m"(values[6]), "=m"(values[7]),
+               "=&r"(v0), "=&r"(v1), "=&r"(v2),
+               "=&r"(v3), "=&r"(v4), "=&r"(v5),
+               "+&r"(code1), "+&r"(code2)
+             : "r"(levels - 1), "r"(-levels), "r"(ff_inverse[levels]));
+
+    return code1 | code2;
+}
+
+#endif
 
 #if HAVE_NEON && HAVE_INLINE_ASM && HAVE_ASM_MOD_Y
 

libavcodec/dca.c

@@ -1038,6 +1038,7 @@ static void dca_downmix(float *samples, int srcfmt,
 }
 
 
+#ifndef decode_blockcodes
 /* Very compact version of the block code decoder that does not use table
  * look-up but is slightly slower */
 static int decode_blockcode(int code, int levels, int *values)
@@ -1051,13 +1052,15 @@ static int decode_blockcode(int code, int levels, int *values)
         code = div;
     }
 
-    if (code == 0)
-        return 0;
-    else {
-        av_log(NULL, AV_LOG_ERROR, "ERROR: block code look-up failed\n");
-        return AVERROR_INVALIDDATA;
-    }
+    return code;
+}
+
+static int decode_blockcodes(int code1, int code2, int levels, int *values)
+{
+    return decode_blockcode(code1, levels, values) |
+           decode_blockcode(code2, levels, values + 4);
 }
+#endif
 
 static const uint8_t abits_sizes[7] = { 7, 10, 12, 13, 15, 17, 19 };
 static const uint8_t abits_levels[7] = { 3, 5, 7, 9, 13, 17, 25 };
@@ -1125,16 +1128,20 @@ static int dca_subsubframe(DCAContext * s, int base_channel, int block_index)
                 if (abits >= 11 || !dca_smpl_bitalloc[abits].vlc[sel].table){
                     if (abits <= 7){
                         /* Block code */
-                        int block_code1, block_code2, size, levels;
+                        int block_code1, block_code2, size, levels, err;
 
                         size = abits_sizes[abits-1];
                         levels = abits_levels[abits-1];
 
                         block_code1 = get_bits(&s->gb, size);
-                        /* FIXME Should test return value */
-                        decode_blockcode(block_code1, levels, block);
                         block_code2 = get_bits(&s->gb, size);
-                        decode_blockcode(block_code2, levels, &block[4]);
+                        err = decode_blockcodes(block_code1, block_code2,
+                                                levels, block);
+                        if (err) {
+                            av_log(s->avctx, AV_LOG_ERROR,
+                                   "ERROR: block code look-up failed\n");
+                            return AVERROR_INVALIDDATA;
+                        }
                     }else{
                         /* no coding */
                         for (m = 0; m < 8; m++)