[arm neon] data type convert (float <-> int32_t)

The conclusion is:

• NEON intrinsics support converting a float data type to int32 regarding a Q value. The reverse is also supported.
• Neon intrinsics already do the clip operation to float value outside of [-1, 1]:!!
• The issue cost 1 Cycles per 32x2 on cortex-a9.

 

Source code:

#include "arm_neon.h"
void test_float_int32_converter()
{
    const int q = 31;
    float   f[] = {-1.0, -0.5, 0, 0.5};
    int32_t i[] = {0, 0, 0, 0};

    float32x4_t vf = vld1q_f32(&f[0]);

    int32x4_t   vi = vcvtq_n_s32_f32(vf, q);
    vst1q_s32(&i[0], vi);

    printf ( "Q=%d\n" , q);
    printf ( "src float 32x4: \t%f,\t%f,\t%f,\t%f\n" , f[0], f[1], f[2], f[3]);
    printf ( "dst int   32x4: \t%8.8x,\t%8.8x,\t%8.8x,\t%8.8x\n" , i[0], i[1], i[2], i[3]);

    vf = vcvtq_n_f32_s32(vi, q);
    vst1q_f32(&f[0], vf);

    printf ( "dst float 32x4: \t%f,\t%f,\t%f,\t%f\n" , f[0], f[1], f[2], f[3]);

    return ;
}

build:

 

armcc --arm_linux_paths --arm_linux_config_file= /home/jxion/arm_linux_config .xml main.c --cpu=Cortex-A9 -o test .exe

 

Result on Pandaboard:

Q=31
src float 32x4:         -1.000000,      -0.500000,      0.000000,       0.500000
dst int   32x4:         80000000,       c0000000,       00000000,       40000000
dst float 32x4:         -1.000000,      -0.500000,      0.000000,       0.500000

Q=31
src float 32x4:         -1.200000,      -0.500000,      1.000000,       1.500000
dst int   32x4:         80000000,       c0000000,       7fffffff,       7fffffff
dst float 32x4:         -1.000000,      -0.500000,      1.000000,       1.000000