Mult plus shift left ops using MMX assembler instructions

Question

I am looking for doing shl(mult(var1,var2),1) operation, where mult multiplies var1 and var2 (both are 16-bit signed integers) and shl shifts left arithmetically the multiplication result. Result must be saturated, i.e., int32 max or int32 min if overflow or underflow occurs and mult(-32768,-32768)=2147483647.

I need to make this op for multiple values in an efficient way for which I think using MMX/SSE instruction set. I though about making mult(sign_extesion(var1), shl(sign_extension(var2))) but I've just discovered no MMX mult() saturation version exists. Do you know any other way to get it?

Answer 1

I think the following should work for you. There is only one potential overflow case (SHRT_MIN * SHRT_MIN) and it handles this explicitly:

#include <limits.h>
#include <mmintrin.h>

int main(void)
{        
    __m64 v1 = _mm_set_pi16(0, SHRT_MAX, 0, SHRT_MIN);
    __m64 v2 = _mm_set_pi16(0, SHRT_MIN, 0, SHRT_MIN);
    __m64 v = _mm_madd_pi16(v1, v2); // 16 x 16 signed multiply
    v = _mm_slli_pi32(v, 1);         // shift left by 1 bit to get full range
    __m64 vcmp = _mm_cmpeq_pi32(v, _mm_set1_pi32(INT_MIN));
                                     // test for SHRT_MIN * SHRT_MIN overflow
    v = _mm_add_pi32(v, vcmp);       // and correct if needed

    return 0;
}