diff options
| author | Thom Johansen <thomj@rockbox.org> | 2007-11-19 16:31:10 +0000 |
|---|---|---|
| committer | Thom Johansen <thomj@rockbox.org> | 2007-11-19 16:31:10 +0000 |
| commit | 3947b0c6328e446fd70641885a7035614cfc90e5 (patch) | |
| tree | 9703371e86781472cd911fbf8079a8cbc2538da6 | |
| parent | 0fc7fb3686c0e52ceac843c1a91f1070f1a70bf6 (diff) | |
FS #8179. Small Musepack speedup for ARM.
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@15692 a1c6a512-1295-4272-9138-f99709370657
| -rw-r--r-- | apps/codecs/libmusepack/synth_filter.c | 79 |
1 files changed, 37 insertions, 42 deletions
diff --git a/apps/codecs/libmusepack/synth_filter.c b/apps/codecs/libmusepack/synth_filter.c index eca23804c5..978e48521b 100644 --- a/apps/codecs/libmusepack/synth_filter.c +++ b/apps/codecs/libmusepack/synth_filter.c @@ -448,50 +448,45 @@ mpc_decoder_windowing_D(MPC_SAMPLE_FORMAT * Data, const MPC_SAMPLE_FORMAT * V) // 32=32x32-multiply assembler for ARM for ( k = 0; k < 32; k++, V++ ) { - asm volatile ( - "ldmia %[D]!, { r0-r3 } \n\t" - "ldr r4, [%[V]] \n\t" - "mul r5, r0, r4 \n\t" - "ldr r4, [%[V], #96*4] \n\t" - "mla r5, r1, r4, r5 \n\t" - "ldr r4, [%[V], #128*4] \n\t" - "mla r5, r2, r4, r5 \n\t" - "ldr r4, [%[V], #224*4] \n\t" - "mla r5, r3, r4, r5 \n\t" - - "ldmia %[D]!, { r0-r3 } \n\t" - "ldr r4, [%[V], #256*4] \n\t" - "mla r5, r0, r4, r5 \n\t" - "ldr r4, [%[V], #352*4] \n\t" - "mla r5, r1, r4, r5 \n\t" - "ldr r4, [%[V], #384*4] \n\t" - "mla r5, r2, r4, r5 \n\t" - "ldr r4, [%[V], #480*4] \n\t" - "mla r5, r3, r4, r5 \n\t" - - "ldmia %[D]!, { r0-r3 } \n\t" - "ldr r4, [%[V], #512*4] \n\t" - "mla r5, r0, r4, r5 \n\t" - "ldr r4, [%[V], #608*4] \n\t" - "mla r5, r1, r4, r5 \n\t" - "ldr r4, [%[V], #640*4] \n\t" - "mla r5, r2, r4, r5 \n\t" - "ldr r4, [%[V], #736*4] \n\t" - "mla r5, r3, r4, r5 \n\t" - - "ldmia %[D]!, { r0-r3 } \n\t" - "ldr r4, [%[V], #768*4] \n\t" - "mla r5, r0, r4, r5 \n\t" - "ldr r4, [%[V], #864*4] \n\t" - "mla r5, r1, r4, r5 \n\t" - "ldr r4, [%[V], #896*4] \n\t" - "mla r5, r2, r4, r5 \n\t" - "ldr r4, [%[V], #992*4] \n\t" - "mla r5, r3, r4, r5 \n\t" - "str r5, [%[Data]], #4 \n" + asm volatile ( + "ldmia %[D]!, { r0-r7 } \n\t" + "ldr r8, [%[V]] \n\t" + "mul r9, r0, r8 \n\t" + "ldr r8, [%[V], #96*4] \n\t" + "mla r9, r1, r8, r9 \n\t" + "ldr r8, [%[V], #128*4] \n\t" + "mla r9, r2, r8, r9 \n\t" + "ldr r8, [%[V], #224*4] \n\t" + "mla r9, r3, r8, r9 \n\t" + "ldr r8, [%[V], #256*4] \n\t" + "mla r9, r4, r8, r9 \n\t" + "ldr r8, [%[V], #352*4] \n\t" + "mla r9, r5, r8, r9 \n\t" + "ldr r8, [%[V], #384*4] \n\t" + "mla r9, r6, r8, r9 \n\t" + "ldr r8, [%[V], #480*4] \n\t" + "mla r9, r7, r8, r9 \n\t" + "ldmia %[D]!, { r0-r7 } \n\t" + "ldr r8, [%[V], #512*4] \n\t" + "mla r9, r0, r8, r9 \n\t" + "ldr r8, [%[V], #608*4] \n\t" + "mla r9, r1, r8, r9 \n\t" + "ldr r8, [%[V], #640*4] \n\t" + "mla r9, r2, r8, r9 \n\t" + "ldr r8, [%[V], #736*4] \n\t" + "mla r9, r3, r8, r9 \n\t" + "ldr r8, [%[V], #768*4] \n\t" + "mla r9, r4, r8, r9 \n\t" + "ldr r8, [%[V], #864*4] \n\t" + "mla r9, r5, r8, r9 \n\t" + "ldr r8, [%[V], #896*4] \n\t" + "mla r9, r6, r8, r9 \n\t" + "ldr r8, [%[V], #992*4] \n\t" + "mla r9, r7, r8, r9 \n\t" + "str r9, [%[Data]], #4 \n" : [Data] "+r" (Data), [D] "+r" (D) : [V] "r" (V) - : "r0", "r1", "r2", "r3", "r4", "r5"); + : "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7", "r8", "r9"); } #else // 32=32x32-multiply (FIXED_POINT) |