typedef short int int16_t;
typedef unsigned char uint8_t;
typedef __builtin_neon_qi int8x8_t __attribute__ ((__vector_size__ (8)));
typedef __builtin_neon_uqi uint8x8_t __attribute__ ((__vector_size__ (8)));
typedef __builtin_neon_hi int16x8_t __attribute__ ((__vector_size__ (16)));
typedef struct int8x8x4_t
{
  int16x8_t val[4];
} int16x8x4_t;
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
vqshrun_n_s16 (int16x8_t __a, const int __b)
{
  return (uint8x8_t)__builtin_neon_vqshrun_nv8hi (__a, __b, 1);
}
__extension__ static __inline void __attribute__ ((__always_inline__))
vst1_u8 (uint8_t * __a, uint8x8_t __b)
{
  __builtin_neon_vst1v8qi ((__builtin_neon_qi *) __a, (int8x8_t) __b);
}
__extension__ static __inline int16x8_t __attribute__ ((__always_inline__))
vqaddq_s16 (int16x8_t __a, int16x8_t __b)
{
  return (int16x8_t)__builtin_neon_vqaddv8hi (__a, __b, 1);
}
__extension__ static __inline int16x8_t __attribute__ ((__always_inline__))
vmulq_s16 (int16x8_t __a, int16x8_t __b)
{
  return (int16x8_t)__builtin_neon_vmulv8hi (__a, __b, 1);
}
__extension__ static __inline int16x8_t __attribute__ ((__always_inline__))
vdupq_n_s16 (int16_t __a)
{
  return (int16x8_t)__builtin_neon_vdup_nv8hi ((__builtin_neon_hi) __a);
}
__extension__ static __inline int16x8x4_t __attribute__ ((__always_inline__))
vld4q_s16 (const int16_t * __a)
{
  union { int16x8x4_t __i; __builtin_neon_xi __o; } __rv;
  __rv.__o = __builtin_neon_vld4v8hi ((const __builtin_neon_hi *) __a);
  return __rv.__i;
}
typedef unsigned char JSAMPLE;
typedef short JCOEF;
typedef unsigned int JDIMENSION;
typedef JSAMPLE *JSAMPROW;
typedef JSAMPROW *JSAMPARRAY;
typedef JCOEF *JCOEFPTR;
typedef struct {
  void * dct_table;
} jpeg_component_info;
typedef struct jpeg_decompress_struct * j_decompress_ptr;
typedef struct {
  int16x8x4_t half[2];
} int16_8x8_t;
void jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
                JCOEFPTR coef_block, JSAMPARRAY output_buf,
                JDIMENSION output_col)
{
  int16_8x8_t coef;
  coef.half[0] = vld4q_s16((int16_t*) coef_block);
  coef.half[1] = vld4q_s16(((int16_t*) coef_block) + 32);
  {
    int16x8x4_t dequant;
    coef.half[0].val[1] = vmulq_s16(coef.half[0].val[1], dequant.val[1]);
    dequant = vld4q_s16(((int16_t*) compptr->dct_table) + 32);
    coef.half[1].val[3] = vmulq_s16(coef.half[1].val[3], dequant.val[3]);
    int16x8_t vector_one_half = vdupq_n_s16(0x80 << 5);
    vst1_u8((uint8_t*)(output_buf[0] + output_col), vqshrun_n_s16(vqaddq_s16(coef.half[0].val[0], vector_one_half), 5));
    vst1_u8((uint8_t*)(output_buf[1] + output_col), vqshrun_n_s16(vqaddq_s16(coef.half[0].val[1], vector_one_half), 5));
    vst1_u8((uint8_t*)(output_buf[7] + output_col), vqshrun_n_s16(vqaddq_s16(coef.half[1].val[3], vector_one_half), 5));
  }
}