Add preliminary ARMv5te optimisations (verified working, but can probably be sped up further), and fix a comment in the ARMv6 code.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@18693 a1c6a512-1295-4272-9138-f99709370657
This commit is contained in:
Jens Arnold 2008-10-03 09:33:36 +00:00
parent 6fcf2765dd
commit 46bf6bd493
3 changed files with 205 additions and 3 deletions

View file

@ -32,6 +32,8 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110, USA
#include "vector_math16_cf.h" #include "vector_math16_cf.h"
#elif ARM_ARCH >= 6 #elif ARM_ARCH >= 6
#include "vector_math16_armv6.h" #include "vector_math16_armv6.h"
#elif ARM_ARCH >= 5 /* Assume all our ARMv5 targets are ARMv5te(j) */
#include "vector_math16_armv5te.h"
#elif defined CPU_ARM7TDMI #elif defined CPU_ARM7TDMI
#include "vector_math16_arm7.h" #include "vector_math16_arm7.h"
#else #else

View file

@ -0,0 +1,202 @@
/*
libdemac - A Monkey's Audio decoder
$Id$
Copyright (C) Dave Chapman 2007
ARMv5te vector math copyright (C) 2008 Jens Arnold
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110, USA
*/
static inline void vector_add(int16_t* v1, int16_t* v2)
{
#if ORDER > 32
int order = (ORDER >> 5);
while (order--)
#endif
{
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
#if ORDER > 16
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
*v1++ += *v2++;
#endif
}
}
static inline void vector_sub(int16_t* v1, int16_t* v2)
{
#if ORDER > 32
int order = (ORDER >> 5);
while (order--)
#endif
{
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
#if ORDER > 16
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
*v1++ -= *v2++;
#endif
}
}
/* This version fetches data as 32 bit words, and *requires* v1 to be
* 32 bit aligned, otherwise it will result either in a data abort, or
* incorrect results (if ARM aligncheck is disabled). */
static inline int32_t scalarproduct(int16_t* v1, int16_t* v2)
{
int res = 0;
#if ORDER > 16
int cnt = ORDER>>4;
#endif
asm volatile (
"tst %[v2], #2 \n"
"beq 20f \n"
"10: \n"
"ldrh r4, [%[v2]], #2 \n"
"mov r4, r4, lsl #16 \n"
"1: \n"
"ldmia %[v1]!, {r0-r3} \n"
"ldmia %[v2]!, {r5-r8} \n"
"smlabt %[res], r0, r4, %[res] \n"
"smlatb %[res], r0, r5, %[res] \n"
"smlabt %[res], r1, r5, %[res] \n"
"smlatb %[res], r1, r6, %[res] \n"
"smlabt %[res], r2, r6, %[res] \n"
"smlatb %[res], r2, r7, %[res] \n"
"smlabt %[res], r3, r7, %[res] \n"
"smlatb %[res], r3, r8, %[res] \n"
"mov r4, r8 \n"
"ldmia %[v1]!, {r0-r3} \n"
"ldmia %[v2]!, {r5-r8} \n"
"smlabt %[res], r0, r4, %[res] \n"
"smlatb %[res], r0, r5, %[res] \n"
"smlabt %[res], r1, r5, %[res] \n"
"smlatb %[res], r1, r6, %[res] \n"
"smlabt %[res], r2, r6, %[res] \n"
"smlatb %[res], r2, r7, %[res] \n"
"smlabt %[res], r3, r7, %[res] \n"
"smlatb %[res], r3, r8, %[res] \n"
#if ORDER > 16
"mov r4, r8 \n"
"subs %[cnt], %[cnt], #1 \n"
"bne 1b \n"
#endif
"b 99f \n"
"20: \n"
"1: \n"
"ldmia %[v1]!, {r0-r3} \n"
"ldmia %[v2]!, {r4-r7} \n"
"smlabb %[res], r0, r4, %[res] \n"
"smlatt %[res], r0, r4, %[res] \n"
"smlabb %[res], r1, r5, %[res] \n"
"smlatt %[res], r1, r5, %[res] \n"
"smlabb %[res], r2, r6, %[res] \n"
"smlatt %[res], r2, r6, %[res] \n"
"smlabb %[res], r3, r7, %[res] \n"
"smlatt %[res], r3, r7, %[res] \n"
"ldmia %[v1]!, {r0-r3} \n"
"ldmia %[v2]!, {r4-r7} \n"
"smlabb %[res], r0, r4, %[res] \n"
"smlatt %[res], r0, r4, %[res] \n"
"smlabb %[res], r1, r5, %[res] \n"
"smlatt %[res], r1, r5, %[res] \n"
"smlabb %[res], r2, r6, %[res] \n"
"smlatt %[res], r2, r6, %[res] \n"
"smlabb %[res], r3, r7, %[res] \n"
"smlatt %[res], r3, r7, %[res] \n"
#if ORDER > 16
"subs %[cnt], %[cnt], #1 \n"
"bne 1b \n"
#endif
"99: \n"
: /* outputs */
#if ORDER > 16
[cnt]"+r"(cnt),
#endif
[v1] "+r"(v1),
[v2] "+r"(v2),
[res]"+r"(res)
: /* inputs */
: /* clobbers */
"r0", "r1", "r2", "r3", "r4",
"r5", "r6", "r7", "r8"
);
return res;
}

View file

@ -205,9 +205,7 @@ static inline void vector_sub(int16_t* v1, int16_t* v2)
/* This version fetches data as 32 bit words, and *requires* v1 to be /* This version fetches data as 32 bit words, and *requires* v1 to be
* 32 bit aligned, otherwise it will result either in a data abort, or * 32 bit aligned, otherwise it will result either in a data abort, or
* incorrect results (if ARM aligncheck is disabled). It is optimised * incorrect results (if ARM aligncheck is disabled). */
* for ARM7TDMI. Using it for ARM9 or higher results in worse performance
* than the C version. */
static inline int32_t scalarproduct(int16_t* v1, int16_t* v2) static inline int32_t scalarproduct(int16_t* v1, int16_t* v2)
{ {
int res = 0; int res = 0;