void test_vaddQs8 (void)
{
int8x16_t out_int8x16_t;
int8x16_t arg0_int8x16_t;
int8x16_t arg1_int8x16_t;
out_int8x16_t = vaddq_s8 (arg0_int8x16_t, arg1_int8x16_t);
}
/* s8x16 add */
void mw_neon_mm_add_s8x16(signed char * A, int Row, int Col, signed char * B, signed char * C)
{
int8x16_t neon_a, neon_b, neon_c;
int size = Row * Col;
int i = 0;
int k = 0;
for (i = 16; i <= size ; i+=16)
{
k = i - 16;
neon_a = vld1q_s8(A + k);
neon_b = vld1q_s8(B + k);
neon_c = vaddq_s8(neon_a, neon_b);
vst1q_s8(C + k, neon_c);
}
k = i - 16;
for (i = 0; i < size % 16; i++)
{
C[k + i] = A[k + i] + B[k + i];
}
}
// Neon multi-vector types are homogeneous aggregates.
// CHECK: define arm_aapcs_vfpcc <16 x i8> @f0(%struct.int8x16x4_t %{{.*}})
// CHECK64: define <16 x i8> @f0(<16 x i8> %{{.*}}, <16 x i8> %{{.*}}, <16 x i8> %{{.*}}, <16 x i8> %{{.*}})
int8x16_t f0(int8x16x4_t v4) {
return vaddq_s8(v4.val[0], v4.val[3]);
}
/* s8x16 mv mul */
void mw_neon_mv_mul_s8x16(signed char * A, int Row, int T, signed char * B, signed char * C)
{
int i = 0;
int k = 0;
int8x16_t neon_b, neon_c;
int8x16_t neon_a0, neon_a1, neon_a2, neon_a3, neon_a4, neon_a5, neon_a6, neon_a7;
int8x16_t neon_a8, neon_a9, neon_a10, neon_a11, neon_a12, neon_a13, neon_a14, neon_a15;
int8x16_t neon_b0, neon_b1, neon_b2, neon_b3, neon_b4, neon_b5, neon_b6, neon_b7;
int8x16_t neon_b8, neon_b9, neon_b10, neon_b11, neon_b12, neon_b13, neon_b14, neon_b15;
for (i = 0; i < Row; i+=16)
{
neon_c = vmovq_n_s8(0);
for (k = 0; k < T; k+=16)
{
int j = k * T + i;
neon_a0 = vld1q_s8(A + j);
j+=Row;
neon_a1 = vld1q_s8(A + j);
j+=Row;
neon_a2 = vld1q_s8(A + j);
j+=Row;
neon_a3 = vld1q_s8(A + j);
j+=Row;
neon_a4 = vld1q_s8(A + j);
j+=Row;
neon_a5 = vld1q_s8(A + j);
j+=Row;
neon_a6 = vld1q_s8(A + j);
j+=Row;
neon_a7 = vld1q_s8(A + j);
j+=Row;
neon_a8 = vld1q_s8(A + j);
j+=Row;
neon_a9 = vld1q_s8(A + j);
j+=Row;
neon_a10 = vld1q_s8(A + j);
j+=Row;
neon_a11 = vld1q_s8(A + j);
j+=Row;
neon_a12 = vld1q_s8(A + j);
j+=Row;
neon_a13 = vld1q_s8(A + j);
j+=Row;
neon_a14 = vld1q_s8(A + j);
j+=Row;
neon_a15 = vld1q_s8(A + j);
neon_b = vld1q_s8(B + k);
neon_b0 = vdupq_n_s8(vgetq_lane_s8(neon_b, 0));
neon_b1 = vdupq_n_s8(vgetq_lane_s8(neon_b, 1));
neon_b2 = vdupq_n_s8(vgetq_lane_s8(neon_b, 2));
neon_b3 = vdupq_n_s8(vgetq_lane_s8(neon_b, 3));
neon_b4 = vdupq_n_s8(vgetq_lane_s8(neon_b, 4));
neon_b5 = vdupq_n_s8(vgetq_lane_s8(neon_b, 5));
neon_b6 = vdupq_n_s8(vgetq_lane_s8(neon_b, 6));
neon_b7 = vdupq_n_s8(vgetq_lane_s8(neon_b, 7));
neon_b8 = vdupq_n_s8(vgetq_lane_s8(neon_b, 8));
neon_b9 = vdupq_n_s8(vgetq_lane_s8(neon_b, 9));
neon_b10 = vdupq_n_s8(vgetq_lane_s8(neon_b, 10));
neon_b11 = vdupq_n_s8(vgetq_lane_s8(neon_b, 11));
neon_b12 = vdupq_n_s8(vgetq_lane_s8(neon_b, 12));
neon_b13 = vdupq_n_s8(vgetq_lane_s8(neon_b, 13));
neon_b14 = vdupq_n_s8(vgetq_lane_s8(neon_b, 14));
neon_b15 = vdupq_n_s8(vgetq_lane_s8(neon_b, 15));
neon_c = vaddq_s8(vmulq_s8(neon_a0, neon_b0), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a1, neon_b1), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a2, neon_b2), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a3, neon_b3), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a4, neon_b4), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a5, neon_b5), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a6, neon_b6), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a7, neon_b7), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a8, neon_b8), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a9, neon_b9), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a10, neon_b10), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a11, neon_b11), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a12, neon_b12), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a13, neon_b13), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a14, neon_b14), neon_c);
neon_c = vaddq_s8(vmulq_s8(neon_a15, neon_b15), neon_c);
}
vst1q_s8(C + i, neon_c);
}
}
inline int8x16_t vaddq(const int8x16_t & v0, const int8x16_t & v1) { return vaddq_s8 (v0, v1); }
