普通卷积armv7-neon指令集实现—QNNPACK

最新推荐文章于 2023-06-29 09:15:37 发布
最新推荐文章于 2023-06-29 09:15:37 发布
阅读量944 收藏 2
点赞数 1
分类专栏： QNNPACK 文章标签： 神经网络
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/qq_20880415/article/details/108432459
版权

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);
  }
}

 

WTHunt
关注
  • 1
    点赞
  • 踩
  • 2
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
专栏目录
【ARMv8/ARMv9 硬件加速系列 2 -- ARM NEON 加速运算介绍】
CodingCos的博客
06-18 206
在ARMv8架构中，引入了一组新的寄存器，称为向量寄存器（Vector Registers），用于支持高效的向量和浮点计算。一个关键的概念是，这些向量寄存器可以以多种不同的视图来访问，而不是仅限于一种固定的数据类型或大小。这意味着同一寄存器中的数据可以被解释为不同位宽的整数或不同精度的浮点数，具体取决于执行的指令类型。ARMv8架构通过这32个128位的向量寄存器，提供了强大的并行计算能力，极大地增强了处理器的数据处理能力，尤其在需要快速处理大规模数据集的现代计算任务中。每个向量寄存器的大小为128位。
ARM Neon并行加速第一课
技术改变生活，创新让世界更美
06-14 3593
转载请标明出处：https://blog.csdn.net/u013752202/article/details/92008192 文章目的： 让小白也能迅速开始Neon编程。 一、什么是Neon Neon是ARMARMv7-A架构以上的处理器（从Cortex-A5开始）中集成的一套SIMD(SingleInstruction,MultipleData）单指令多数据指令集，相当于X...
ARM NEON 使用手册
09-17
ARM NEON是什么东西我就不多做介绍了，我觉得想用这技术的多半是高手，高手一般都有CSDN下载分的！哈哈哈！如果没有下载分，私信我留下邮箱，我会发给你。具体看链接：http://www.arm.com/zh/products/processors/technologies/neon.php
ARM NEON 编程系列3——使用ARM NEON Intrinsics加速Video Codec
STN_LCD的专栏
08-18 3602
http://www.jianshu.com/p/70601b36540f 前言 最近公司在视频直播项目中要使用H.265/HEVC，具体的是使用HW硬件编码H.264/AVC，云端转码成H.265/HEVC并推流的解决方案。方案中使用的解码器是FFMpeg中的H.265解码器，该解码器是从OpenHEVC直接获取的，比起备受好评的H.264/AVC解码器，这个解码器目
使用NEON指令加速浮点数转整数
老猫的专栏
08-25 3620
使用NEON指令的好处就是一次可以转4个float到int，但是有个不好处理的事情就是四舍五入。使用 vcvtq_s32_f32是可以转f32到s32，但它是截断的，不会做四舍五入处理。这时候有个小窍门就是对正的浮点数加上0.5f，对负的浮点数减去0.5f，再调用vcvtq_s32_f32就相当于是四舍五入了。 float ftmp[] = {0.5f, -1.0f, 0.49999997f...
armneon Intrinsics记录
邹九的个人博客
01-08 872
vld3q_f16将float16的数组，依次导入到3个寄存器，想RGBRGBRGBRGBRGB......的数据，就是R值一个寄存器，G值一个寄存器，B值一个寄存器，每个寄存器放8个，占用128bit，返回类型为float16x8x3_t，指令里带了q的都是Q单元寄存器存在128bit vld1_u8是将uint8类型的数组，存放到1个寄存器内，每个寄存器放8个，占用64bit，返回类型为uint8x8_t指令里不带q的，是D单元寄存器存在64bit float16x8x3_t，从右向左，就是3个
性能优化篇（3）：NEON快速入门指南
woquhan的博客
11-27 8621
本文介绍了一些 NEON Intrinsics，涉及：向量赋值、访问向量、存储向量以及向量算术逻辑运算。
cudnn-6.5-linux-ARMv7-v2.tgz
06-27
它包括一系列高效的卷积神经网络（CNN）、递归神经网络（RNN）和长短期记忆网络（LSTM）的实现，这些在网络模型中非常常见。通过使用CuDNN，开发者可以快速构建和训练复杂的深度学习模型，而不必关注底层硬件的细节...
cudnn-6.5-linux-armv7-R1.tgz
06-27
CUDNN（CUDA Deep Neural Network）是CUDA的一个核心库，专门为深度学习算法设计，提供了高度优化的GPU实现，用于卷积神经网络（CNN）、递归神经网络（RNN）和其他深度学习模型。 描述简单明了，表明这是一个可供...
cudnn-7.0-android-armv7-v3.0-prod.tgz
06-27
针对ARMv7架构，这是Android设备常用的处理器架构，意味着此版本的CUDNN是为那些使用ARMv7指令集的移动设备设计的，比如一些旧款或中低端的智能手机和平板电脑。CUDNN的Android版本使得开发者可以在这些设备上运行...
cudnn-7.0-android-ARMv7-v4.0-prod.tgz
最新发布
06-30
在Android ARMv7环境下使用CuDNN，开发者需要注意性能调优，包括选择最适合当前任务的卷积算法、优化内存分配和减少数据传输开销。此外，考虑到移动设备的资源限制，还需要关注模型的大小和计算复杂度，以实现最佳的...
cudnn-7.0-linux-armv7-v3.0-prod.tgz
06-27
CUDNN（CUDA Deep Neural Network）是NVIDIA开发的一个深度学习库，它针对卷积神经网络、递归神经网络等进行了高度优化，提供了对GPU加速的深度学习算法的接口。CUDNN-7.0是这一库的一个版本，主要服务于深度学习...
NEON指令集释义
独行侠
09-24 9162
目录 知识直通车 常用NEON指令集 neon指令总结 知识直通车 neon指令权威官网：https://developer.arm.com/architectures/instruction-sets/simd-isas/neon/intrinsics neon指令中文速查：https://blog.csdn.net/billbliss/article/details/7892463...
常用NEON 内置函数记录备用
热门推荐
VivienFu的专栏
12-15 2万+
float32x4_t _sum1 = vld1q_f32(outptr);//load outptr地址起始的4个float数据到_sum1 float32x4_t _sum2 = vdupq_n_f32(0.f);//声明_sum2,存储的四个float32都初始化为0 float32x4_t _r02 = vextq_f32(_r00, _r00n, 2)
NEON学习笔记（2）NEON Intrinsics
s_trive_的博客
06-29 1075
ARM 编译器工具链中的 NEON Intrinsics是一种编写 NEON 代码的方法，该代码比汇编程序更容易维护，同时仍然保持对生成哪些 NEON 指令的控制。NEON Intrinsics是编译器用适当的 NEON 指令或 NEON 指令序列替换的函数调用。然而，您必须优化代码以充分利用 NEON 单元提供的速度提升。有些数据类型对应于包含不同大小元素的 NEON 寄存器（D 寄存器和 Q 寄存器）。这些允许您创建直接映射到 NEON 寄存器的 C 变量。这些变量被传递给 NEON 内部函数。
ARM NEON编程初探——一个简单的BGR888转YUV444实例详解
weixin_34161083的博客
07-10 937
原文请猛戳这里 最近在学习ARM的SIMD指令集NEON，发现这方面的资料真是太少了，我便来给NEON凑凑人气，姑且以这篇入门文章来分享一些心得吧。 学习一门新技术，总是有一些经典是绕不开的，对于NEON来说，这份必备的武林秘籍自然就是ARM官方的《NEON Programmer's Guide》（以下简称Guide）啦。别看这份Gui...
高性能计算学习笔记（四）
weixin_47034794的博客
11-09 664
高性能计算学习笔记（四）IPP软件移植简单实践举例
基于NEON指令的图像旋转加速【armv7】
独行侠
10-27 3750
目录 前言 知识直通车 NEON转置指令 右旋90 4x4矩阵右旋实例 灰度图（单通道）右旋90 彩图（RGB三通道）右旋90 左旋90 4x4矩阵左旋实例 灰度图（单通道）左旋90 彩图（RGB三通道）左旋90 RK3288-linux(硬浮点)耗时对比 前言 图像旋转就是简单的坐标变换，1080p的图像右旋90度耗时时间，在RK3288-linux(硬浮点)板子...
ARM SIMD 指令集：NEON 简介
Man
03-12 1万+
ARM SIMD指令集——NEON简介SIMD简介NEON简介指令向量数据类型内在函数参考连接 SIMD简介 SIMD，即 single instruction multiple data，单指令流多数据流，也就是说一次运算指令可以执行多个数据流，从而提高程序的运算速度。 NEON简介 NEON 是一种压缩的 SIMD 架构，由ARMv7引入，在ARMv8对其功能进行了扩展，支持包括加法、乘法、比较、移位、绝对值 、极大极小极值运算、保存和加载指令等运算。 ARMV7-A/R ARMV8-
卷积网络Lenet-5的优点
04-29
卷积网络LeNet-5是一个经典的卷积神经网络结构，它的优点包括： 1. 参数共享：卷积操作可以使得每个神经元的参数被共享，从而降低了模型的参数数量，减少了过拟合的风险。 2. 局部连接：卷积操作只考虑输入数据的...

“相关推荐”对你有帮助么？

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值