ISPC 8-bit/16-bit

gistfile1.txt

export void simple(uniform unsigned int8 vin[], uniform unsigned int8 vout[], uniform int count) {
    foreach (index = 0 ... count) {
        unsigned int8 v = vin[index];
        // Do some calculations in 16-bit lanes. I expect v to be split in 2 registers
        unsigned int16 v2 = ((unsigned int16)v) * ((unsigned int16)v);
        // And somehow fold back into 8-bit lanes
        unsigned int8 m = v2>>8;
        // Write result
        vout[index] = m;
    }
}

Assembly for full body is along the lines of (for "--target=avx1-i32x8"):

	pmovzxbw	(%rcx,%rax), %xmm1 # xmm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
	pmullw	%xmm1, %xmm1
	psrlw	$8, %xmm1
	pshufb	%xmm0, %xmm1
	movq	%xmm1, (%rdx,%rax)
	addl	$8, %eax
	cmpl	%r9d, %eax

Maybe this is faster, not sure. But easier to prototype with intrinsics in this case.
Note that with other targets the code might get better, but I didn't manage to make it so.