diff options
| -rwxr-xr-x | build_iphone.sh (renamed from iphone.sh) | 14 | ||||
| -rwxr-xr-x | configure | 1 | ||||
| -rw-r--r-- | configure.ac | 1 | ||||
| -rw-r--r-- | src/macros.h | 12 | ||||
| -rw-r--r-- | src/neon.s | 58 |
5 files changed, 71 insertions, 15 deletions
diff --git a/iphone.sh b/build_iphone.sh index c970c92..70d7845 100755 --- a/iphone.sh +++ b/build_iphone.sh @@ -1,8 +1,18 @@ #/bin/sh +INSTALL_DIR="`pwd`/build" + export SDKVER="6.0" export DEVROOT="/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer" export SDKROOT="$DEVROOT/SDKs/iPhoneOS$SDKVER.sdk" -export CFLAGS="-O3 -Wreturn-type -Wparentheses -Wswitch -Wno-unused-parameter -Wno-unused-variable -Wunused-value -Wno-shorten-64-to-32 -Wno-trigraphs -fpascal-strings -miphoneos-version-min=5.0 -mcpu=cortex-a9 -arch armv7 -mfpu=neon -pipe -isysroot $SDKROOT -isystem $SDKROOT/usr/include -isystem $DEVROOT/usr/include -std=c11 -mno-thumb" +export CFLAGS="-O3 -Wreturn-type -Wparentheses -Wswitch -Wno-unused-parameter -Wno-unused-variable -Wunused-value -Wno-shorten-64-to-32 -Wno-trigraphs -fpascal-strings -miphoneos-version-min=5.0 -mcpu=cortex-a9 -arch armv7 -mfpu=neon -pipe -isysroot $SDKROOT -isystem $SDKROOT/usr/include -isystem $DEVROOT/usr/include -std=c99 -mno-thumb -no-integrated-as" export AR="$DEVROOT/usr/bin/ar" -#export CC="$DEVROOT/usr/bin/clang" +export CC="clang" + + +mkdir -p $INSTALL_DIR +./configure --enable-neon --build=i386-apple-darwin10.8.0 --host=arm-eabi --prefix=$INSTALL_DIR + +make +make install + @@ -15999,6 +15999,7 @@ $as_echo "no, using $LN_S" >&6; } fi + # Checks for library functions. #AC_FUNC_MALLOC for ac_func in gettimeofday pow diff --git a/configure.ac b/configure.ac index dfd476f..2b5de20 100644 --- a/configure.ac +++ b/configure.ac @@ -76,6 +76,7 @@ AC_PROG_CC_STDC AC_PROG_INSTALL AC_PROG_LN_S AC_PROG_LIBTOOL + # Checks for library functions. #AC_FUNC_MALLOC AC_CHECK_FUNCS([gettimeofday pow]) diff --git a/src/macros.h b/src/macros.h index 63464f4..ade5f42 100644 --- a/src/macros.h +++ b/src/macros.h @@ -48,10 +48,14 @@ static inline V VLIT4(data_t f3, data_t f2, data_t f1, data_t f0) { #define FFTS_FREE(d) (free(d)) __INLINE void STORESPR(data_t * addr, VS p) { - __asm__ __volatile__ ("vst1.32 {%q1,%q2}, [%0, :128]\n\t" - : - : "r" (addr), "w" (p.val[0]), "w" (p.val[1]) - : "memory"); + + vst1q_f32(addr, p.val[0]); + vst1q_f32(addr + 4, p.val[1]); + +//__asm__ __volatile__ ("vst1.32 {%q1,%q2}, [%0, :128]\n\t" +// : +// : "r" (addr), "w" (p.val[0]), "w" (p.val[1]) +// : "memory"); } #else @@ -1,7 +1,13 @@ +#include "asm.s" - .globl neon_x4 .align 4 +#ifdef __APPLE__ + .globl _neon_x4 +_neon_x4: +#else + .globl neon_x4 neon_x4: +#endif @ add r3, r0, #0 vld1.32 {q8,q9}, [r0, :128] @@ -43,9 +49,14 @@ neon_x4: vst1.32 {q6,q7}, [r6, :128] bx lr - .globl neon_x8 .align 4 +#ifdef __APPLE__ + .globl _neon_x8 +_neon_x8: +#else + .globl neon_x8 neon_x8: +#endif mov r11, #0 add r3, r0, #0 @ data0 add r5, r0, r1, lsl #1 @ data2 @@ -159,9 +170,14 @@ neon_x8_loop: bx lr - .globl neon_x8_t .align 4 +#ifdef __APPLE__ + .globl _neon_x8_t +_neon_x8_t: +#else + .globl neon_x8_t neon_x8_t: +#endif mov r11, #0 add r3, r0, #0 @ data0 add r5, r0, r1, lsl #1 @ data2 @@ -281,9 +297,14 @@ neon_x8_t_loop: @ r3-r10 = data pointers @ r11 = loop iterations @ r2 & lr = temps - .globl neon_ee .align 4 +#ifdef __APPLE__ + .globl _neon_ee +_neon_ee: +#else + .globl neon_ee neon_ee: +#endif vld1.32 {d16, d17}, [r2, :128] _neon_ee_loop: vld2.32 {q15}, [r10, :128]! @@ -357,10 +378,14 @@ _neon_ee_loop: @ r3-r10 = data pointers @ r11 = loop iterations @ r2 & lr = temps - .globl neon_oo .align 4 +#ifdef __APPLE__ + .globl _neon_oo +_neon_oo: +#else + .globl neon_oo neon_oo: - +#endif _neon_oo_loop: vld2.32 {q8}, [r6, :128]! vld2.32 {q9}, [r5, :128]! @@ -411,9 +436,14 @@ _neon_oo_loop: @ r3-r10 = data pointers @ r11 = addr of twiddle @ r2 & lr = temps - .globl neon_eo .align 4 +#ifdef __APPLE__ + .globl _neon_eo +_neon_eo: +#else + .globl neon_eo neon_eo: +#endif vld2.32 {q9}, [r5, :128]! @tag2 vld2.32 {q13}, [r3, :128]! @tag0 vld2.32 {q12}, [r4, :128]! @tag1 @@ -488,9 +518,14 @@ neon_eo: @ r3-r10 = data pointers @ r11 = addr of twiddle @ r2 & lr = temps - .globl neon_oe .align 4 +#ifdef __APPLE__ + .globl _neon_oe +_neon_oe: +#else + .globl neon_oe neon_oe: +#endif vld1.32 {q8}, [r5, :128]! vld1.32 {q10}, [r6, :128]! vld2.32 {q11}, [r4, :128]! @@ -564,7 +599,12 @@ neon_oe: vstmia lr!, {q4-q7} + .align 4 +#ifdef __APPLE__ + .globl _neon_end +_neon_end: +#else .globl neon_end - .align 4 neon_end: +#endif bx lr |
