Arm64 Assembly Language

Modern Arm Assembly Language Programming: Covers Armv8-A 32-bit, 64-bit, and SIMD

by Daniel Kusswurm

2021.07.28: updated by

Chapter 15: Armv8-64 SIMD Floating-Point Programming

Packed Floating-Point Arithmetic

Comparisons

NEON を使って比較演算を行うプログラムの説明。

ch15_02/main.cpp

#include <iostream>
#include <iomanip>
#include <cmath>
#include "Vec128.h"
using namespace std;

extern void PackedCompareF32_(Vec128 x[8], const Vec128& a, const Vec128& b);
extern void PackedCompareF64_(Vec128 x[8], const Vec128& a, const Vec128& b);

const char* c_CmpStr[8] = {
    "EQ", "NE", "GT", "GE", "LT", "LE", "a LT0", "b GT0"
};

void PackedCompareF32(void) {
    const char nl = '\n';
    Vec128 x[8], a, b;

    a.m_F32[0] = 2.0;         b.m_F32[0] = -4.0;
    a.m_F32[1] = 17.0;        b.m_F32[1] = 12.0;
    a.m_F32[2] = -6.0;        b.m_F32[2] = -6.0;
    a.m_F32[3] = 3.0;         b.m_F32[3] = 8.0;

    PackedCompareF32_(x, a, b);

    cout << "\nResults for PackedCompareF32_\n";
    cout << setw(11) << 'a' << ':' << a.ToStringF32() << nl;
    cout << setw(11) << 'b' << ':' << b.ToStringF32() << nl;
    cout << nl;

    for (int j = 0; j < 8; j++)
        cout << setw(11) << c_CmpStr[j] << ':' << x[j].ToStringX32() << nl;
}

void PackedCompareF64(void)
{
    const char nl = '\n';
    Vec128 x[8], a, b;

    a.m_F64[0] = -2.0;        b.m_F64[0] = -4.0;
    a.m_F64[1] = M_SQRT2;     b.m_F64[1] = M_PI;

    PackedCompareF64_(x, a, b);

    cout << "\nResults for PackedCompareF64_\n";
    cout << setw(11) << 'a' << ':' << a.ToStringF64() << nl;
    cout << setw(11) << 'b' << ':' << b.ToStringF64() << nl;
    cout << nl;

    for (int j = 0; j < 8; j++)
        cout << setw(11) << c_CmpStr[j] << ':' << x[j].ToStringX64() << nl;
}

int main()
{
    PackedCompareF32();
    PackedCompareF64();
    return 0;
}

ch15_02/neon.cpp

#include "Vec128.h"

void PackedCompareF32_(Vec128 x[8], const Vec128& a, const Vec128& b) {
  __asm volatile("\n\
	ld1	{v0.4s}, [x1]            // v0 = a           \n\
	ld1	{v1.4s}, [x2]            // v1 = b           \n\
	fcmeq	v2.4s, v0.4s, v1.4s      // packed a == b    \n\
	st1	{v2.4s}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	not     v2.16b, v2.16b           // packed a !=b     \n\
	st1	{v2.4s}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmgt	v2.4s, v0.4s, v1.4s      // packed a > b     \n\
	st1	{v2.4s}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmge	v2.4s, v0.4s, v1.4s      // packed a >= b    \n\
	st1	{v2.4s}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmlt	v2.4s, v0.4s, v1.4s      // packed a < b     \n\
	st1	{v2.4s}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmle	v2.4s, v0.4s, v1.4s      // packed a <= b    \n\
	st1	{v2.4s}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmlt	v2.4s, v0.4s, 0.0        // packed a < 0     \n\
	st1	{v2.4s}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmgt	v2.4s, v1.4s, 0.0        // packed b > 0     \n\
	st1	{v2.4s}, [x0], 16        // [x0]=v2; x0+=16  \n\
"
		 :
		 :
		 : "v0", "v1", "v2"
		 );
}

void PackedCompareF64_(Vec128 x[8], const Vec128& a, const Vec128& b) {
  __asm volatile("\n\
	ld1	{v0.2d}, [x1]            // v0 = a           \n\
	ld1	{v1.2d}, [x2]            // v1 = b           \n\
	fcmeq	v2.2d, v0.2d, v1.2d      // packed a == b    \n\
	st1	{v2.2d}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	not     v2.16b, v2.16b           // packed a !=b     \n\
	st1	{v2.2d}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmgt	v2.2d, v0.2d, v1.2d      // packed a > b     \n\
	st1	{v2.2d}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmge	v2.2d, v0.2d, v1.2d      // packed a >= b    \n\
	st1	{v2.2d}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmlt	v2.2d, v0.2d, v1.2d      // packed a < b     \n\
	st1	{v2.2d}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmle	v2.2d, v0.2d, v1.2d      // packed a <= b    \n\
	st1	{v2.2d}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmlt	v2.2d, v0.2d, 0.0        // packed a < 0     \n\
	st1	{v2.2d}, [x0], 16        // [x0]=v2; x0+=16  \n\
                                                             \n\
	fcmgt	v2.2d, v1.2d, 0.0        // packed b > 0     \n\
	st1	{v2.2d}, [x0], 16        // [x0]=v2; x0+=16  \n\
"
		 :
		 :
		 : "v0", "v1", "v2"
		 );
}

ch15_02/main.cpp の実行例

arm64@manet Ch15_02 % g++ -I.. -std=c++11 -O main.cpp neon.cpp -o a.out
arm64@manet Ch15_02 % ./a.out

Results for PackedCompareF32_
          a:        2.000000       17.000000  |       -6.000000        3.000000
          b:       -4.000000       12.000000  |       -6.000000        8.000000

         EQ:        00000000        00000000  |        FFFFFFFF        00000000
         NE:        FFFFFFFF        FFFFFFFF  |        00000000        FFFFFFFF
         GT:        FFFFFFFF        FFFFFFFF  |        00000000        00000000
         GE:        FFFFFFFF        FFFFFFFF  |        FFFFFFFF        00000000
         LT:        00000000        00000000  |        00000000        FFFFFFFF
         LE:        00000000        00000000  |        FFFFFFFF        FFFFFFFF
      a LT0:        00000000        00000000  |        FFFFFFFF        00000000
      b GT0:        00000000        FFFFFFFF  |        00000000        FFFFFFFF

Results for PackedCompareF64_
          a:                 -2.000000000000  |                  1.414213562373
          b:                 -4.000000000000  |                  3.141592653590

         EQ:                0000000000000000  |                0000000000000000
         NE:                FFFFFFFFFFFFFFFF  |                FFFFFFFFFFFFFFFF
         GT:                FFFFFFFFFFFFFFFF  |                0000000000000000
         GE:                FFFFFFFFFFFFFFFF  |                0000000000000000
         LT:                0000000000000000  |                FFFFFFFFFFFFFFFF
         LE:                0000000000000000  |                FFFFFFFFFFFFFFFF
      a LT0:                FFFFFFFFFFFFFFFF  |                0000000000000000
      b GT0:                0000000000000000  |                FFFFFFFFFFFFFFFF

http://nw.tsuda.ac.jp/