登录
当前位置： 首页 > 工具软件 > QNNPACK > 使用案例 >

普通卷积armv7-neon指令集实现—QNNPACK

琴献
2023-12-01

与浮点计算一样，依旧是4x8的分块，为防止饱和将uint8塞进int16进行计算，这样一个寄存器装载8个数，本篇主要关注指令集。

指令集总结：

vld1_dup_u8(const uint8_t*):广播broadcast uint8_t到8x8

vld1_u8(const uint8_t*)：加载8x8

vmovl_u8(const uint8x8_t):扩展指令集，返回uint16x8

vreinterpretq_s16_u16(const uint16x8_t):类型转换

vsubl_u8(const uint8x8_t,const uint8x8_t):减完返回16x8

int32x4_t = vmlal_lane_s16(int32x4_t, int16x4_t, int16x4_t, int);

vget_low_s16(const int16x8_t):取低int16x4_t

vget_high_s16(const int16x8_t):取高int16x4_t

 代码：

void q8conv_ukernel_4x8__neon(
    size_t mr,
    size_t nr,
    size_t kc,
    size_t ks,
    const uint8_t** a,
    const void* w,
    int32_t* c,
    size_t c_stride,
    const union qnnp_conv_quantization_params* quantization_params)
{
  const uint8x8_t vb_zero_point = vld1_dup_u8((const uint8_t*) &quantization_params->neon.kernel_zero_point);

  int32x4_t vacc0x0123 = vld1q_s32((int32_t*)w); w = (void*) ((uintptr_t) w + sizeof(int32x4_t));
  int32x4_t vacc0x4567 = vld1q_s32((int32_t*)w); w = (void*) ((uintptr_t) w + sizeof(int32x4_t));
  int32x4_t vacc1x0123 = vacc0x0123;
  int32x4_t vacc1x4567 = vacc0x4567;
  int32x4_t vacc2x0123 = vacc0x0123;
  int32x4_t vacc2x4567 = vacc0x4567;
  int32x4_t vacc3x0123 = vacc0x0123;
  int32x4_t vacc3x4567 = vacc0x4567;

  do {
    const uint8_t* a0 = *a++;
    const uint8_t* a1 = *a++;
    const uint8_t* a2 = *a++;
    const uint8_t* a3 = *a++;

    size_t k = kc;
    for (; k >= 8; k -= 8) {
      const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
      const uint8x8_t va1 = vld1_u8(a1); a1 += 8;
      const uint8x8_t va2 = vld1_u8(a2); a2 += 8;
      const uint8x8_t va3 = vld1_u8(a3); a3 += 8;
      const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
      const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
      const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
      const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 0);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 0);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 0);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 0);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 0);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 0);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 0);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 0);
      }

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 1);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 1);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 1);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 1);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 1);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 1);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 1);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 1);
      }

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 2);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 2);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 2);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 2);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 2);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 2);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 2);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 2);
      }

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 3);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 3);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 3);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 3);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 3);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 3);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 3);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 3);
      }

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 0);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 0);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 0);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 0);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 0);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 0);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 0);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 0);
      }

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 1);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 1);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 1);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 1);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 1);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 1);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 1);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 1);
      }

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 2);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 2);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 2);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 2);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 2);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 2);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 2);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 2);
      }

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 3);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 3);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 3);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 3);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 3);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 3);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 3);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 3);
      }
    }
    if (k != 0) {
      const size_t a_predecrement = 8 - k;
      const int64x1_t va_shift = vmov_n_s64(-8 * a_predecrement);
      const uint8x8_t va0 = vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(a0 - a_predecrement)), va_shift));
      const uint8x8_t va1 = vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(a1 - a_predecrement)), va_shift));
      const uint8x8_t va2 = vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(a2 - a_predecrement)), va_shift));
      const uint8x8_t va3 = vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(a3 - a_predecrement)), va_shift));
      const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
      const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
      const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
      const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));

      {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 0);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 0);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 0);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 0);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 0);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 0);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 0);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 0);
      }

      if (k >= 2) {
        const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
        const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

        vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 1);
        vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 1);
        vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 1);
        vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 1);
        vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 1);
        vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 1);
        vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 1);
        vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 1);

        if (k > 2) {
          const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 2);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 2);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 2);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 2);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 2);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 2);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 2);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 2);

          if (k >= 4) {
            const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
            const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

            vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 3);
            vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 3);
            vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 3);
            vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 3);
            vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 3);
            vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 3);
            vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 3);
            vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 3);

            if (k > 4) {
              const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
              const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

              vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 0);
              vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 0);
              vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 0);
              vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 0);
              vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 0);
              vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 0);
              vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 0);
              vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 0);

              if (k >= 6) {
                const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
                const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

                vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 1);
                vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 1);
                vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 1);
                vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 1);
                vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 1);
                vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 1);
                vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 1);
                vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 1);

                if (k > 6) {
                  const uint8x8_t vb01234567 = vld1_u8((uint8_t*)w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
                  const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

                  vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 2);
                  vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 2);
                  vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 2);
                  vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 2);
                  vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 2);
                  vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 2);
                  vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 2);
                  vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 2);
                }
              }
            }
          }
        }
      }
    }
  } while (--ks != 0);

  int32_t* c0 = c;
  int32_t* c1 = c0 + c_stride;
  if (mr < 2) {
    c1 = c0;
  }
  int32_t* c2 = c1 + c_stride;
  if (mr <= 2) {
    c2 = c1;
  }
  int32_t* c3 = c2 + c_stride;
  if (mr != 4) {
    c3 = c2;
  }

  int32_t tempResult[4];
  if (nr == 8) {
    vst1q_s32(c0,vacc0x0123); vst1q_s32(c0+4,vacc0x4567);
    vst1q_s32(c1,vacc1x0123); vst1q_s32(c1+4,vacc1x4567);
    vst1q_s32(c2,vacc2x0123); vst1q_s32(c2+4,vacc2x4567);
    vst1q_s32(c3,vacc3x0123); vst1q_s32(c3+4,vacc3x4567);
  } else if(nr > 4) {
      vst1q_s32(c0,vacc0x0123); 
      vst1q_s32(tempResult,vacc0x4567); memcpy(c0+4,tempResult,sizeof(int32_t)*(nr-4));

      vst1q_s32(c1,vacc1x0123); 
      vst1q_s32(tempResult,vacc1x4567); memcpy(c1+4,tempResult,sizeof(int32_t)*(nr-4));

      vst1q_s32(c2,vacc2x0123); 
      vst1q_s32(tempResult,vacc2x4567); memcpy(c2+4,tempResult,sizeof(int32_t)*(nr-4));

      vst1q_s32(c3,vacc3x0123); 
      vst1q_s32(tempResult,vacc3x4567); memcpy(c3+4,tempResult,sizeof(int32_t)*(nr-4));
  } else if(nr != 0){
      vst1q_s32(tempResult,vacc0x0123); memcpy(c0,tempResult,sizeof(int32_t)*nr);
      vst1q_s32(tempResult,vacc1x0123); memcpy(c1,tempResult,sizeof(int32_t)*nr);
      vst1q_s32(tempResult,vacc2x0123); memcpy(c2,tempResult,sizeof(int32_t)*nr);
      vst1q_s32(tempResult,vacc3x0123); memcpy(c3,tempResult,sizeof(int32_t)*nr);
  }
}

 

 类似资料：

相关阅读

Pytorch 实现sobel算子的卷积操作详解Tensorflow实现卷积神经网络的详细代码PyTorch上实现卷积神经网络CNN的方法什么是卷积？Android使用Notification实现普通通知栏（一）

相关文章

AngularJS 指令JSP指令Pycharm微指令JSP taglib指令JSP include指令

相关问答

卷积输出卷积网络的输入通道数如何在运行时检测NEON和氦指令集的可用性用Java8实现集合的笛卡尔积Cortex-A57能否双重发布128位neon指令？

相关文档

Java 集合学习指南Xcode Server 持续集成指南Phodal's Idea 实战指南Burp Suite 实战指南Python 最佳实践指南
快捷导航： 新手教程 算法原理 架构设计 Java进阶 数据库进阶 大厂专栏 面试经验 编程笔记 编程问答 所有专题 文档资料 工具软件 电子书籍 小牛导航
在线工具： 房贷计算器 个税计算器 Linux命令查询 Json格式化 正则表达式 颜色转换 AES加解密 SHA1加密 MD5加密 毒鸡汤 字数统计 随机密码生成 进制转换 Base64编解码 励志句子
Copyright © 2019-2024 小牛知识库@xnip.cn. All Rights Reserved.