typedef short int int16_t;
typedef int int32_t;
typedef long int int64_t;
typedef enum pa_sample_format {
	PA_SAMPLE_U8, PA_SAMPLE_ALAW, PA_SAMPLE_ULAW, PA_SAMPLE_S16LE,
	    PA_SAMPLE_S16BE, PA_SAMPLE_FLOAT32LE, PA_SAMPLE_FLOAT32BE,
	    PA_SAMPLE_S32LE, PA_SAMPLE_S32BE, PA_SAMPLE_S24LE, PA_SAMPLE_S24BE,
	    PA_SAMPLE_S24_32LE, PA_SAMPLE_S24_32BE, PA_SAMPLE_MAX,
	    PA_SAMPLE_INVALID = -1
} pa_log_target_t;
typedef enum pa_cpu_x86_flag {
	PA_CPU_X86_MMX = (1 << 0), PA_CPU_X86_MMXEXT =
	    (1 << 1), PA_CPU_X86_SSE = (1 << 2), PA_CPU_X86_SSE2 =
	    (1 << 3), PA_CPU_X86_SSE3 = (1 << 4), PA_CPU_X86_SSSE3 =
	    (1 << 5), PA_CPU_X86_SSE4_1 = (1 << 6), PA_CPU_X86_SSE4_2 =
	    (1 << 7), PA_CPU_X86_3DNOW = (1 << 8), PA_CPU_X86_3DNOWEXT =
	    (1 << 9)
} pa_cpu_x86_flag_t;
typedef int64_t pa_reg_x86;
typedef void (*pa_do_volume_func_t) (void *samples, void *volumes,
				     unsigned channels, unsigned length);

void pa_volume_s16ne_mmx(short * samples, int * volumes,
				unsigned channels, unsigned length)
{
	long int channel;
	long int temp;
	__asm__ __volatile__(" xor %3, %3                    \n\t"
			     " sar $1, %2                    \n\t"
			     " test $1, %2                   \n\t"
			     " je 2f                         \n\t"
			     " movd (%1, %3, 4), %%mm0       \n\t"
			     " movw (%0), %w4                \n\t"
			     " movd %4, %%mm1                \n\t"
			     " pxor  %%mm4, %%mm4           \n\t"
			     " punpcklwd %%mm4, " "%%mm1" "        \n\t"
			     " pcmpgtw " "%%mm0" ", %%mm4          \n\t"
			     " pand " "%%mm1" ", %%mm4             \n\t"
			     " movq " "%%mm1" ", %%mm5             \n\t"
			     " pmulhw " "%%mm0" ", " "%%mm1" "            \n\t"
			     " paddw %%mm4, " "%%mm1" "            \n\t"
			     " psrld $16, " "%%mm0" "              \n\t"
			     " pmaddwd %%mm5, " "%%mm0" "          \n\t"
			     " paddd " "%%mm1" ", " "%%mm0" "             \n\t"
			     " packssdw " "%%mm0" ", " "%%mm0" "          \n\t"
			     " movd %%mm0, %4                \n\t"
			     " movw %w4, (%0)                \n\t"
			     " add $2, %0                    \n\t" " add " "$1"
			     ", %3                 \n\t"
			     " mov %3, %4                   \n\t" " sub " "%5"
			     ", %4                 \n\t"
			     " cmovae %4, %3                \n\t"
			     "2:                             \n\t"
			     " sar $1, %2                    \n\t"
			     " test $1, %2                   \n\t"
			     " je 4f                         \n\t"
			     "3:                             \n\t"
			     " movq (%1, %3, 4), %%mm0       \n\t"
			     " movd (%0), %%mm1              \n\t"
			     " pxor  %%mm4, %%mm4           \n\t"
			     " punpcklwd %%mm4, " "%%mm1" "        \n\t"
			     " pcmpgtw " "%%mm0" ", %%mm4          \n\t"
			     " pand " "%%mm1" ", %%mm4             \n\t"
			     " movq " "%%mm1" ", %%mm5             \n\t"
			     " pmulhw " "%%mm0" ", " "%%mm1" "            \n\t"
			     " paddw %%mm4, " "%%mm1" "            \n\t"
			     " psrld $16, " "%%mm0" "              \n\t"
			     " pmaddwd %%mm5, " "%%mm0" "          \n\t"
			     " paddd " "%%mm1" ", " "%%mm0" "             \n\t"
			     " packssdw " "%%mm0" ", " "%%mm0" "          \n\t"
			     " movd %%mm0, (%0)              \n\t"
			     " add $4, %0                    \n\t" " add " "$2"
			     ", %3                 \n\t"
			     " mov %3, %4                   \n\t" " sub " "%5"
			     ", %4                 \n\t"
			     " cmovae %4, %3                \n\t"
			     "4:                             \n\t"
			     " sar $1, %2                    \n\t"
			     " cmp $0, %2                    \n\t"
			     " je 6f                         \n\t"
			     "5:                             \n\t"
			     " movq (%1, %3, 4), %%mm0       \n\t"
			     " movq 8(%1, %3, 4), %%mm2      \n\t"
			     " movd (%0), %%mm1              \n\t"
			     " movd 4(%0), %%mm3             \n\t"
			     " pxor  %%mm4, %%mm4           \n\t"
			     " punpcklwd %%mm4, " "%%mm1" "        \n\t"
			     " pcmpgtw " "%%mm0" ", %%mm4          \n\t"
			     " pand " "%%mm1" ", %%mm4             \n\t"
			     " movq " "%%mm1" ", %%mm5             \n\t"
			     " pmulhw " "%%mm0" ", " "%%mm1" "            \n\t"
			     " paddw %%mm4, " "%%mm1" "            \n\t"
			     " psrld $16, " "%%mm0" "              \n\t"
			     " pmaddwd %%mm5, " "%%mm0" "          \n\t"
			     " paddd " "%%mm1" ", " "%%mm0" "             \n\t"
			     " packssdw " "%%mm0" ", " "%%mm0" "          \n\t"
			     " pxor  %%mm4, %%mm4           \n\t"
			     " punpcklwd %%mm4, " "%%mm3" "        \n\t"
			     " pcmpgtw " "%%mm2" ", %%mm4          \n\t"
			     " pand " "%%mm3" ", %%mm4             \n\t"
			     " movq " "%%mm3" ", %%mm5             \n\t"
			     " pmulhw " "%%mm2" ", " "%%mm3" "            \n\t"
			     " paddw %%mm4, " "%%mm3" "            \n\t"
			     " psrld $16, " "%%mm2" "              \n\t"
			     " pmaddwd %%mm5, " "%%mm2" "          \n\t"
			     " paddd " "%%mm3" ", " "%%mm2" "             \n\t"
			     " packssdw " "%%mm2" ", " "%%mm2" "          \n\t"
			     " movd %%mm0, (%0)              \n\t"
			     " movd %%mm2, 4(%0)             \n\t"
			     " add $8, %0                    \n\t" " add " "$4"
			     ", %3                 \n\t"
			     " mov %3, %4                   \n\t" " sub " "%5"
			     ", %4                 \n\t"
			     " cmovae %4, %3                \n\t"
			     " dec %2                        \n\t"
			     " jne 5b                        \n\t"
			     "6:                             \n\t"
			     " emms                          \n\t":"+r"
			     (samples), "+r"(volumes), "+r"(length),
			     "=D"((pa_reg_x86) channel),
			     "=&r"(temp):"X"((pa_reg_x86) channels):"cc");
}

void pa_volume_func_init_mmx(pa_cpu_x86_flag_t flags)
{
	if (flags & PA_CPU_X86_MMX) {
		pa_set_volume_func(PA_SAMPLE_S16LE,
				   (pa_do_volume_func_t) pa_volume_s16ne_mmx);
	}
}