Thanks for your reply.
Here is an example:
#include<arm_neon.h>
void neon_convert (uint8_t * __restrict dest, uint8_t * __restrict src, int n)
{
int i;
uint8x8_t rfac = vdup_n_u8 (77);
uint8x8_t gfac = vdup_n_u8 (151);
uint8x8_t bfac = vdup_n_u8 (28);
n/=8;
for (i=0; i<n; i++)
{
uint16x8_t temp;
uint8x8x3_t rgb = vld3_u8 (src);
uint8x8x3_t rgb1 = vld3_u8 (src+2);
uint8x8_t result;
temp = vmull_u8 (rgb1.val[0], rfac);
temp = vmlal_u8 (temp,rgb.val[1], gfac);
temp = vmlal_u8 (temp,rgb.val[2], bfac);
result = vshrn_n_u16 (temp, 8);
vst1_u8 (dest, result);
src += 8*3;
dest += 8;
}
}
The above code on compilation produces the following output on objdump:
Disassembly of section .text.neon_convert:
00000000 <neon_convert>:
0: e92d 0ff0 stmdb sp!, {r4, r5, r6, r7, r8, r9, sl, fp}
4: f102 0a07 add.w sl, r2, #7 ; 0x7
8: ea32 0222 bics.w r2, r2, r2, asr #32
c: bf28 it cs
e: 4652 movcs r2, sl
10: f04f 034d mov.w r3, #77 ; 0x4d
14: eec5 3b90 vdup.8 d21, r3
18: f06f 0368 mvn.w r3, #104 ; 0x68
1c: ea4f 0ae2 mov.w sl, r2, asr #3
20: eec4 3b90 vdup.8 d20, r3
24: f1ba 0f00 cmp.w sl, #0 ; 0x0
28: f04f 031c mov.w r3, #28 ; 0x1c
2c: b09a sub sp, #104
2e: 4604 mov r4, r0
30: 468c mov ip, r1
32: eec3 3b90 vdup.8 d19, r3
36: dd3b ble.n b0 <neon_convert+0xb0>
38: 2500 movs r5, #0
3a: af08 add r7, sp, #32
3c: f10d 0950 add.w r9, sp, #80 ; 0x50
40: f10d 0808 add.w r8, sp, #8 ; 0x8
44: f10d 0b38 add.w fp, sp, #56 ; 0x38
48: f96c 040f vld3.8 {d16-d18}, [ip]
4c: 463e mov r6, r7
4e: ecc7 0b06 vstmia r7, {d16-d18}
52: 3501 adds r5, #1
54: ce0f ldmia r6!, {r0, r1, r2, r3}
56: 9601 str r6, [sp, #4]
58: 464e mov r6, r9
5a: c60f stmia r6!, {r0, r1, r2, r3}
5c: 9a01 ldr r2, [sp, #4]
5e: f10c 0302 add.w r3, ip, #2 ; 0x2
62: f10c 0c18 add.w ip, ip, #24 ; 0x18
66: e892 0003 ldmia.w r2, {r0, r1}
6a: e886 0003 stmia.w r6, {r0, r1}
6e: f963 040f vld3.8 {d16-d18}, [r3]
72: ecc8 0b06 vstmia r8, {d16-d18}
76: 4646 mov r6, r8
78: eddd 2b16 vldr d18, [sp, #88]
7c: ce0f ldmia r6!, {r0, r1, r2, r3}
7e: 9601 str r6, [sp, #4]
80: 465e mov r6, fp
82: c60f stmia r6!, {r0, r1, r2, r3}
84: eddd 0b0e vldr d16, [sp, #56]
88: 9a01 ldr r2, [sp, #4]
8a: ffc0 0ca5 vmull.u8 q8, d16, d21
8e: ffc2 08a4 vmlal.u8 q8, d18, d20
92: eddd 2b18 vldr d18, [sp, #96]
96: e892 0003 ldmia.w r2, {r0, r1}
9a: ffc2 08a3 vmlal.u8 q8, d18, d19
9e: efc8 0830 vshrn.i16 d16, q8, #8
a2: e886 0003 stmia.w r6, {r0, r1}
a6: f944 070f vst1.8 {d16}, [r4]
aa: 3408 adds r4, #8
ac: 45aa cmp sl, r5
ae: dccb bgt.n 48 <neon_convert+0x48>
b0: b01a add sp, #104
b2: e8bd 0ff0 ldmia.w sp!, {r4, r5, r6, r7, r8, r9, sl, fp}
b6: 4770 bx lr
On Fri, Jul 16, 2010 at 8:18 PM, Richard Earnshaw <rearnsha@xxxxxxx> wrote:
>
>
> On Fri, 2010-07-09 at 16:59 +0530, Ajmal Ahammed wrote:
> > Sir,
> >
> > My name is Ajmal A. I was compiling a program with neon
> > intrinsics for arm cortex A-8. When I examined the output code, I
> > observed that, among the available 32 d-registers of neon, the program
> > uses only those from d16 (either if it's used as double word registers
> > or quad word registers). And another problem is with the vld
> > instructions. For vld instruction, it always uses the registers d16,
> > d17, .... Even if there are free registers like d26,d27, etc it pushes
> > the contents of the registers d16, d17,... to the stack.
> >
> >
>
> Without a testcase it is impossible to tell for sure what is happening.
> Please post an example.
>
>
> > I am using gcc 4.4.0, compiled using android-ndk-r4.
> >
> > For compiling I used the following command:
> > /home/master/ajmal/ndk/build/
> > prebuilt/linux-x86/arm-eabi-4.4.0/bin/arm-eabi-gcc
> > -I/home/master/ajmal/ndk/build/platforms/android-3/arch-arm/usr/include
> > -fpic -mthumb-interwork -ffunction-sections -funwind-tables
> > -fstack-protector -fno-short-enums -D__ARM_ARCH_5__ -D__ARM_ARCH_5T__
> > -D__ARM_ARCH_5E__ -D__ARM_ARCH_5TE__ -Wno-psabi -march=armv7-a
> > -mfloat-abi=softfp -mthumb -Os -fomit-frame-pointer
> > -fno-strict-aliasing -finline-limit=64 -mfpu=neon
> > -I/home/master/ajmal/ndk/sources/cpufeatures
> > -I/home/master/ajmal/ndk/samples/hello-neon/jni -DANDROID
> > -DHAVE_NEON=1 -Wa,--noexecstack -O3 -DNDEBUG -g -c -MMD -MP -MF
> > /home/master/ajmal/ndk/samples/hello-neon/bin/ndk/local/armeabi-v7a/objs/helloneon/$1.o.d
> > /home/master/ajmal/ndk/samples/hello-neon/jni/$1.c -o
> > /home/master/ajmal/ndk/samples/hello-neon/bin/ndk/local/armeabi-v7a/objs/helloneon/$1.o
> >
> > where $1 is the filename without the extension.
> >
> > I have tried the latest version of gcc (4.5.0) also, with no better results.
> >
> > Can you please tell me if there is a special reason for not using the
> > other d registers, or for the stack operations. I have gone through
> > the gcc manuals and some files gcc sources, but I couldn't clarify the
> > doubt.
> >
> > (My language proficiency is not so good, and this is the first time I
> > am writing an e-mail for such an authority. I am sorry for any
> > mistakes due to them).
> >
> > Thank you in advance.
> > Ajmal. A
>
