There were two kinds of improvements: providing the 3-argument operators (i.e. uint128_add($to, $a, $b)) and optimizing the SV to int128_t conversions.

I think I got the first part handled ok - but I haven't given consideration to the second.
For the record, my "rough estimation script" (for multiplication only) is as follows - in accordance with the usual mantra that I follow when creating objects:

package Sis::UInt128;

use warnings;
use strict;
use Math::GMPz qw(:mpz);
use Benchmark qw(:all);

use Inline C => Config =>
    BUILD_NOISY => 1,
    TYPEMAPS => ['./typemap_128'],
    USING => 'ParseRegExp';


use Inline C => <<'EOC';

SV * create() {
     __uint128 *ps;
     SV * obj_ref, * obj;

     New(42, ps, 1, __uint128);
     if(ps == NULL) croak("Failed to allocate memory in create functio
+n");

     obj_ref = newSV(0);
     obj = newSVrv(obj_ref, "Sis::UInt128");

     sv_setiv(obj, INT2PTR(IV,ps));
     SvREADONLY_on(obj);
     return obj_ref;

}

void _assign(__uint128 * rop, SV * h, SV * l) {
     __uint128 __div = 9223372036854775808ULL;
     unsigned __int64 high, low;

     high = (unsigned __int64)SvUV(h);
     low  = (unsigned __int64)SvUV(l);
     *rop = ((__uint128)__div * high) + (__uint128)low;
}


void _deref_obj(__uint128 * obj) {
     dXSARGS;

     __uint128 __div = 9223372036854775808ULL;

     ST(0) = sv_2mortal(newSVuv(*obj / __div));
     ST(1) = sv_2mortal(newSVuv(*obj % __div));

     XSRETURN(2);     
}

void DESTROY(__uint128 * obj) {
    printf("Cleaning up\n");
    Safefree(obj);
}

void mul_128(__uint128 * rop, __uint128 * op1, __uint128 * op2) {
     *rop = *op1 * *op2; 
}


EOC

our $mpz1 = Math::GMPz->new('676469752303423489');
our $mpz2 = Math::GMPz->new('776469752999423489');
our $mpz_ret = Math::GMPz::Rmpz_init2(128);

our $i_ret = create();
our $i_1 = create();
our $i_2 = create();

assign($i_1, "$mpz1");
assign($i_2, "$mpz2");

our $count = 50000;

timethese($count * 19, {
'mul_128' => 'mul_128($i_ret, $i_1, $i_2)',
'gmpz' => 'Math::GMPz::Rmpz_mul($mpz_ret, $mpz1, $mpz2)',
});

print retrieve($i_1),"\n", retrieve($i_2),"\n", retrieve($i_ret), "\n"
+;
print "$mpz1\n$mpz2\n$mpz_ret\n";

sub assign {
    my $obj = shift;
    my $num = shift;
    my @args;

    my $a0 = Math::GMPz->new($num) / 9223372036854775808;
    my $a1 = Math::GMPz->new($num) % 9223372036854775808;

    $args[0] = "$a0" + 0;
    $args[1] = "$a1" + 0;

    _assign($obj, $args[0], $args[1]);
}

sub retrieve {
    my @in = _deref_obj($_[0]);
    my $num = Math::GMPz->new($in[0]);
    $num *= 9223372036854775808;
    $num += $in[1];

    return "$num";
}
[download]

with a "typemap_128" that looks like this:

__uint128 *        UI128

INPUT
UI128
            $var = INT2PTR($type, SvIV(SvRV($arg)))
[download]

Cheers,
Rob