Assembly Audio Mixing

gistfile1.c

/**
 * Represents 20ms of 8khz signed linear 16-bit audio
 *
 * This is how much audio gets put in each RTP packet.
 */
typedef int16_t frame_t[160];

/**
 * Fast mix two audio frames together on x86 and x86_64
 *
 * I take advantage of the sse2 instruction PADDSW that uses a 128-bit
 * register to perform saturated addition on eight 16-bit numbers at the same
 * time. I'm about two orders of a magnitude faster than c code.
 */
static void mix160(frame_t to, const frame_t from)
{
    const int blocks = 160 / 8;
    asm volatile (
        "moar: movdqu (%[to]), %%xmm0;"
        "      movdqu (%[from]), %%xmm1;"
        "      paddsw %%xmm1, %%xmm0;"
        "      movdqu %%xmm0, (%[to]);"
        "      add    $16, %[to];"
        "      add    $16, %[from];"
        "      dec    %[blocks];"
        "      cmp    $0, %[blocks];"
        "      jg     moar;"
        : /* no outputs */
        : [to] "r"(to),
          [from] "r"(from),
          [blocks] "r"(blocks)
    );
}