void dotProd_i16_neon(const float *dataf, const float *weightsf, float *vals, const int n, const int len, const float *istd) {
const int16_t *data = (const int16_t *)dataf;
const int16_t *weights = (const int16_t *)weightsf;
weightsf += n * len / 2; // sizeof(float) / sizeof(int16_t)
for (int i = 0; i < n; i += 4) {
int32x4_t accum0 = { 0, 0, 0, 0 };
int32x4_t accum1 = accum0;
int32x4_t accum2 = accum0;
int32x4_t accum3 = accum0;
for (int j = 0; j < len; j += 8) {
int16x4x2_t d0 = vld2_s16(data + j);
int16x4x2_t w0 = vld2_s16(weights);
int16x4x2_t w1 = vld2_s16(weights + 8);
int16x4x2_t w2 = vld2_s16(weights + 16);
int16x4x2_t w3 = vld2_s16(weights + 24);
accum0 = vmlal_s16(accum0, d0.val[0], w0.val[0]);
accum0 = vmlal_s16(accum0, d0.val[1], w0.val[1]);
accum1 = vmlal_s16(accum1, d0.val[0], w1.val[0]);
accum1 = vmlal_s16(accum1, d0.val[1], w1.val[1]);
accum2 = vmlal_s16(accum2, d0.val[0], w2.val[0]);
accum2 = vmlal_s16(accum2, d0.val[1], w2.val[1]);
accum3 = vmlal_s16(accum3, d0.val[0], w3.val[0]);
accum3 = vmlal_s16(accum3, d0.val[1], w3.val[1]);
weights += 32;
}
int32x2_t sum0 = vpadd_s32(vget_low_s32(accum0), vget_high_s32(accum0));
int32x2_t sum1 = vpadd_s32(vget_low_s32(accum1), vget_high_s32(accum1));
int32x2_t sum2 = vpadd_s32(vget_low_s32(accum2), vget_high_s32(accum2));
int32x2_t sum3 = vpadd_s32(vget_low_s32(accum3), vget_high_s32(accum3));
sum0 = vpadd_s32(sum0, sum1);
sum1 = vpadd_s32(sum2, sum3);
int32x4_t sum = vcombine_s32(sum0, sum1);
float32x4_t val = vcvtq_f32_s32(sum);
val = vmulq_f32(val, vld1q_f32(weightsf + i*2));
val = vmulq_n_f32(val, istd[0]);
val = vaddq_f32(val, vld1q_f32(weightsf + i*2 + 4));
vst1q_f32(vals + i, val);
}
}
void ne10_img_hresize_4channels_linear_neon (const unsigned char** src, int** dst, int count,
const int* xofs, const short* alpha,
int swidth, int dwidth, int cn, int xmin, int xmax)
{
int dx, k;
int dx0 = 0;
int16x4x2_t alpha_vec;
uint8x8_t dS0_vec, dS1_vec;
int16x8_t qS0_vec, qS1_vec;
int16x4_t dS0_0123, dS0_4567, dS1_0123, dS1_4567;
int32x4_t qT0_vec, qT1_vec;
int16x4_t dCoeff;
dCoeff = vdup_n_s16 (INTER_RESIZE_COEF_SCALE);
for (k = 0; k <= count - 2; k++)
{
const unsigned char *S0 = src[k], *S1 = src[k + 1];
int *D0 = dst[k], *D1 = dst[k + 1];
for (dx = dx0; dx < xmax; dx += 4)
{
int sx = xofs[dx];
alpha_vec = vld2_s16 (&alpha[dx * 2]);
dS0_vec = vld1_u8 (&S0[sx]);
dS1_vec = vld1_u8 (&S1[sx]);
qS0_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS0_vec));
qS1_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS1_vec));
dS0_0123 = vget_low_s16 (qS0_vec);
dS0_4567 = vget_high_s16 (qS0_vec);
dS1_0123 = vget_low_s16 (qS1_vec);
dS1_4567 = vget_high_s16 (qS1_vec);
qT0_vec = vmull_s16 (dS0_0123, alpha_vec.val[0]);
qT1_vec = vmull_s16 (dS1_0123, alpha_vec.val[0]);
qT0_vec = vmlal_s16 (qT0_vec, dS0_4567, alpha_vec.val[1]);
qT1_vec = vmlal_s16 (qT1_vec, dS1_4567, alpha_vec.val[1]);
vst1q_s32 (&D0[dx], qT0_vec);
vst1q_s32 (&D1[dx], qT1_vec);
}
for (; dx < dwidth; dx += 4)
{
int sx = xofs[dx];
dS0_vec = vld1_u8 (&S0[sx]);
dS1_vec = vld1_u8 (&S1[sx]);
qS0_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS0_vec));
qS1_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS1_vec));
dS0_0123 = vget_low_s16 (qS0_vec);
dS1_0123 = vget_low_s16 (qS1_vec);
qT0_vec = vmull_s16 (dS0_0123, dCoeff);
qT1_vec = vmull_s16 (dS1_0123, dCoeff);
vst1q_s32 (&D0[dx], qT0_vec);
vst1q_s32 (&D1[dx], qT1_vec);
}
}
for (; k < count; k++)
{
const unsigned char *S = src[k];
int *D = dst[k];
for (dx = 0; dx < xmax; dx += 4)
{
int sx = xofs[dx];
alpha_vec = vld2_s16 (&alpha[dx * 2]);
dS0_vec = vld1_u8 (&S[sx]);
qS0_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS0_vec));
dS0_0123 = vget_low_s16 (qS0_vec);
dS0_4567 = vget_high_s16 (qS0_vec);
qT0_vec = vmull_s16 (dS0_0123, alpha_vec.val[0]);
qT0_vec = vmlal_s16 (qT0_vec, dS0_4567, alpha_vec.val[1]);
vst1q_s32 (&D[dx], qT0_vec);
}
for (; dx < dwidth; dx += 4)
{
int sx = xofs[dx];
dS0_vec = vld1_u8 (&S[sx]);
qS0_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS0_vec));
dS0_0123 = vget_low_s16 (qS0_vec);
qT0_vec = vmull_s16 (dS0_0123, dCoeff);
vst1q_s32 (&D[dx], qT0_vec);
}
}
}
void computeNetwork0new_neon(const float *dataf, const float *weightsf, uint8_t *d) {
const int16_t *data = (const int16_t *)dataf;
const int16_t *weights = (const int16_t *)weightsf;
int32x4_t accum0 = { 0, 0, 0, 0 };
int32x4_t accum1 = accum0;
int32x4_t accum2 = accum0;
int32x4_t accum3 = accum0;
for (int i = 0; i < 128/2; i += 8) {
int16x4x2_t d0 = vld2_s16(data + i);
int16x4x2_t w0 = vld2_s16(weights + i * 4);
int16x4x2_t w1 = vld2_s16(weights + i * 4 + 8);
int16x4x2_t w2 = vld2_s16(weights + i * 4 + 16);
int16x4x2_t w3 = vld2_s16(weights + i * 4 + 24);
accum0 = vmlal_s16(accum0, d0.val[0], w0.val[0]);
accum0 = vmlal_s16(accum0, d0.val[1], w0.val[1]);
accum1 = vmlal_s16(accum1, d0.val[0], w1.val[0]);
accum1 = vmlal_s16(accum1, d0.val[1], w1.val[1]);
accum2 = vmlal_s16(accum2, d0.val[0], w2.val[0]);
accum2 = vmlal_s16(accum2, d0.val[1], w2.val[1]);
accum3 = vmlal_s16(accum3, d0.val[0], w3.val[0]);
accum3 = vmlal_s16(accum3, d0.val[1], w3.val[1]);
}
int32x2_t sum0 = vpadd_s32(vget_low_s32(accum0), vget_high_s32(accum0));
int32x2_t sum1 = vpadd_s32(vget_low_s32(accum1), vget_high_s32(accum1));
int32x2_t sum2 = vpadd_s32(vget_low_s32(accum2), vget_high_s32(accum2));
int32x2_t sum3 = vpadd_s32(vget_low_s32(accum3), vget_high_s32(accum3));
sum0 = vpadd_s32(sum0, sum1);
sum1 = vpadd_s32(sum2, sum3);
int32x4_t sum = vcombine_s32(sum0, sum1);
float32x4_t m0 = vcvtq_f32_s32(sum);
m0 = vmulq_f32(m0, vld1q_f32(weightsf + 512/4));
m0 = vaddq_f32(m0, vld1q_f32(weightsf + 528/4));
float32x4_t m1, m2, m3, m4;
m1 = m0;
m0 = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(m0), sign_bits_f));
m0 = vaddq_f32(m0, ones_f);
m0 = vmulq_f32(reciprocal(m0), m1);
m1 = vdupq_lane_f32(vget_low_f32(m0), 0);
m2 = vdupq_lane_f32(vget_low_f32(m0), 1);
m3 = vdupq_lane_f32(vget_high_f32(m0), 0);
m4 = vdupq_lane_f32(vget_high_f32(m0), 1);
m1 = vmulq_f32(m1, vld1q_f32(weightsf + 544/4));
m2 = vmulq_f32(m2, vld1q_f32(weightsf + 560/4));
m3 = vmulq_f32(m3, vld1q_f32(weightsf + 576/4));
m4 = vmulq_f32(m4, vld1q_f32(weightsf + 592/4));
m1 = vaddq_f32(m1, m2);
m3 = vaddq_f32(m3, m4);
m1 = vaddq_f32(m1, m3);
m1 = vaddq_f32(m1, vld1q_f32(weightsf + 608/4));
uint32x4_t gte = vcgeq_f32(m1, zeroes_f);
uint16x4_t gte_u16 = vmovn_u32(gte);
uint8x8_t gte_u8 = vmovn_u16(vcombine_u16(gte_u16, vget_low_u16(vreinterpretq_u16_u32(sign_bits_f))));
gte_u8 = vshr_n_u8(gte_u8, 7);
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(gte_u8), 0);
}
void computeNetwork0_i16_neon(const float *inputf, const float *weightsf, uint8_t *d) {
const int16_t *input = (const int16_t *)inputf;
const int16_t *weights = (const int16_t *)weightsf;
int32x4_t accum0 = { 0, 0, 0, 0 };
int32x4_t accum1 = accum0;
int32x4_t accum2 = accum0;
int32x4_t accum3 = accum0;
for (int i = 0; i < 96/2; i += 8) {
int16x4x2_t d0 = vld2_s16(input + i);
int16x4x2_t w0 = vld2_s16(weights + i * 4);
int16x4x2_t w1 = vld2_s16(weights + i * 4 + 8);
int16x4x2_t w2 = vld2_s16(weights + i * 4 + 16);
int16x4x2_t w3 = vld2_s16(weights + i * 4 + 24);
accum0 = vmlal_s16(accum0, d0.val[0], w0.val[0]);
accum0 = vmlal_s16(accum0, d0.val[1], w0.val[1]);
accum1 = vmlal_s16(accum1, d0.val[0], w1.val[0]);
accum1 = vmlal_s16(accum1, d0.val[1], w1.val[1]);
accum2 = vmlal_s16(accum2, d0.val[0], w2.val[0]);
accum2 = vmlal_s16(accum2, d0.val[1], w2.val[1]);
accum3 = vmlal_s16(accum3, d0.val[0], w3.val[0]);
accum3 = vmlal_s16(accum3, d0.val[1], w3.val[1]);
}
int32x2_t sum0 = vpadd_s32(vget_low_s32(accum0), vget_high_s32(accum0));
int32x2_t sum1 = vpadd_s32(vget_low_s32(accum1), vget_high_s32(accum1));
int32x2_t sum2 = vpadd_s32(vget_low_s32(accum2), vget_high_s32(accum2));
int32x2_t sum3 = vpadd_s32(vget_low_s32(accum3), vget_high_s32(accum3));
sum0 = vpadd_s32(sum0, sum1);
sum1 = vpadd_s32(sum2, sum3);
int32x4_t sum = vcombine_s32(sum0, sum1);
float32x4_t m0 = vcvtq_f32_s32(sum);
m0 = vmulq_f32(m0, vld1q_f32(weightsf + 384/4));
m0 = vaddq_f32(m0, vld1q_f32(weightsf + 400/4));
float32x4_t m1, m2, m3, m4, m5, m6, m7;
m1 = m0;
m0 = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(m0), sign_bits_f_zero_l));
m0 = vaddq_f32(m0, ones_f);
m0 = vmulq_f32(reciprocal(m0), m1);
m1 = vdupq_lane_f32(vget_low_f32(m0), 0);
m2 = vdupq_lane_f32(vget_low_f32(m0), 1);
m3 = vdupq_lane_f32(vget_high_f32(m0), 0);
m4 = vdupq_lane_f32(vget_high_f32(m0), 1);
m1 = vmulq_f32(m1, vld1q_f32(weightsf + 416/4));
m2 = vmulq_f32(m2, vld1q_f32(weightsf + (416+16)/4));
m3 = vmulq_f32(m3, vld1q_f32(weightsf + (416+32)/4));
m4 = vmulq_f32(m4, vld1q_f32(weightsf + (416+48)/4));
m1 = vaddq_f32(m1, m2);
m3 = vaddq_f32(m3, m4);
m1 = vaddq_f32(m1, m3);
m1 = vaddq_f32(m1, vld1q_f32(weightsf + (416+64)/4));
m7 = m1;
m1 = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(m1), sign_bits_f));
m1 = vaddq_f32(m1, ones_f);
m7 = vmulq_f32(reciprocal(m1), m7);
m3 = m0;
m0 = vdupq_lane_f32(vget_low_f32(m0), 0);
m1 = vdupq_lane_f32(vget_low_f32(m3), 1);
m2 = vdupq_lane_f32(vget_high_f32(m3), 0);
m3 = vdupq_lane_f32(vget_high_f32(m3), 1);
m0 = vmulq_f32(m0, vld1q_f32(weightsf + 496/4));
m1 = vmulq_f32(m1, vld1q_f32(weightsf + (496+16)/4));
m2 = vmulq_f32(m2, vld1q_f32(weightsf + (496+32)/4));
m3 = vmulq_f32(m3, vld1q_f32(weightsf + (496+48)/4));
m4 = vdupq_lane_f32(vget_low_f32(m7), 0);
m5 = vdupq_lane_f32(vget_low_f32(m7), 1);
m6 = vdupq_lane_f32(vget_high_f32(m7), 0);
m7 = vdupq_lane_f32(vget_high_f32(m7), 1);
m4 = vmulq_f32(m4, vld1q_f32(weightsf + (496+64)/4));
m5 = vmulq_f32(m5, vld1q_f32(weightsf + (496+80)/4));
m6 = vmulq_f32(m6, vld1q_f32(weightsf + (496+96)/4));
m7 = vmulq_f32(m7, vld1q_f32(weightsf + (496+112)/4));
m0 = vaddq_f32(m0, m1);
m2 = vaddq_f32(m2, m3);
m4 = vaddq_f32(m4, m5);
m6 = vaddq_f32(m6, m7);
m0 = vaddq_f32(m0, m2);
m4 = vaddq_f32(m4, m6);
m0 = vaddq_f32(m0, m4);
m0 = vaddq_f32(m0, vld1q_f32(weightsf + (496+128)/4));
float32x2_t maximum = vmax_f32(vget_low_f32(m0), vget_high_f32(m0));
d[0] = (vget_lane_f32(maximum, 1) <= vget_lane_f32(maximum, 0));
}
inline int16x4x2_t vld2(const s16 * ptr) { return vld2_s16(ptr); }
void test_vld2s16 (void)
{
int16x4x2_t out_int16x4x2_t;
out_int16x4x2_t = vld2_s16 (0);
}
