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8x8vld1_u8(const uint8_t*)：加载8x8vmovl_u8(const uint8x8_t):扩展指令集，返回uint16x8vreinterpretq_s16_u16(const uint16x8_t):类型转换vsubl_u8(const uint8x8_t,const uint8x8_t):减完返回16x8int32x4_t = vmlal_lane_s16(int32x4_t, int16x4_t, int16x4_t, int);vget_low_s16(const int16x8_t):取低int16x4_tvget_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);}
}

普通卷积armv7-neon指令集实现—QNNPACK相关推荐

1. ARM NEON指令集优化理论与实践

   ARM NEON指令集优化理论与实践 一．简介 NEON就是一种基于SIMD思想的ARM技术,相比于ARMv6或之前的架构,NEON结合了64-bit和128-bit的SIMD指令集,提供128-bi ...

2. neon浮点运算_ARM NEON指令集优化理论与实践

   ARM NEON指令集优化理论与实践 一．简介 NEON就是一种基于SIMD思想的ARM技术,相比于ARMv6或之前的架构,NEON结合了64-bit和128-bit的SIMD指令集,提供128-bi ...

3. 【genius_platform软件平台开发】第二十八讲：NEON指令集优化（附实例）

   当在ARM芯片上进行一些例如图像处理等计算的时候,常常会因为计算量太大造成计算帧率较低的情况.因而,需要选择一种更加简单快捷的计算方式以获得处理速度上的提升.ARM NEON就是一个不错的选择. ※ ...

4. neon 指令 c语言,Neon指令集优化快速入门教程

   Neon指令集优化快速入门教程 Neon指令集优化快速入门教程 文章目录 1. Neon是什么? 2.Neon为什么速度快 3. Neon基础概念 4. Neon的C语言接口 C语言程序中集成Neon ...

5. ARM和neon指令集

   在移动平台上进行一些复杂算法的开发,一般需要用到指令集来进行加速.目前在移动上使用最多的是ARM芯片. ARM是微处理器行业的一家知名企业,其芯片结构有:armv5.armv6.armv7和armv8 ...

6. 【整理】SIMD、MMX、SSE、AVX、3D Now！、neon——指令集大全

   http://blog.csdn.net/conowen/article/details/7255920 SIMD SIMD单指令流多数据流(SingleInstruction Multiple Da ...

7. 论ARMv7 Thumb-2指令集的性能（含Thumb指令集介绍）【转载】

   [摘要] 主要是介绍ARM CPU中的THUMB-2功能,相对于THUMB的比较THUMB-2指令集的扩展,THUMB-2的新指令带来的好处,新指令对性能和代码密度的改进.    如今的嵌入式系统开发 ...

8. 【genius_platform软件平台开发】第八十二讲：ARM Neon指令集一（ARM NEON Intrinsics, SIMD运算, 优化心得）

   1. ARM Neon Intrinsics 编程 1.入门:基本能上手写Intrinsics 1.1 Neon介绍.简明案例与编程惯例 1.2 如何检索Intrinsics 1.3 优化效果案例 1 ...

9. arm 饱和指令_ARM平台下NEON使用方法详解

   NEON介绍 在移动平台上进行一些复杂算法的开发,一般需要用到指令集来进行加速.NEON 技术是 ARM Cortex™-A 系列处理器的 128 位 SIMD(单指令,多数据)架构扩展,专门针对大规 ...

最新文章

1. 实对称矩阵的性质_浅谈矩阵的相似对角化（一）
2. MySQL数据库以及其Python用法
3. 机器学习 KNN算法_0_丐版_鸢尾花集分类(matlab实现)
4. Windows7下用VirtualBox安装Ubuntu网卡配置
5. 无插件，无com组件，利用EXCEL、WORD模板做数据导出（一）
6. mongodb索引使用
7. 杭电3790最短路径问题
8. 18家机构批量刷新SOTA！T5 is all you need！
9. 人工免疫算法与物流中心选址问题
10. 一篇文章完全搞懂正则化（Regularization）
11. php设置北京时间（时区）
12. stc12c5a60s2c语言程序,STC12C5A60S2单片机 花样流水灯（小鸟归巢）C程序
13. iOS开发：国际化之app支持多种语言切换
14. 为什么千万不要小瞧月薪几千的女生
15. yum一次性下载安装包及其依赖包
16. 使用DecimalFormat进行数字的格式化
17. 可以Ping通，却无法使用IE上网的几种可能性
18. Arduino对57步进电机的控制
19. php5.4环境搭建_Windows2003下php5.4安装配置教程（Apache2.4）
20. 【趣拿洞察】奈雪的茶成长背后的商业逻辑

热门文章

1. SharePoint2010 修改模板页在网页中间显示网站内容
2. python_day6 shutil模块
3. border-radius圆角兼容方案
4. ASP.NET MVC网站学习问题积累（一）
5. mysql 暴力破解 root账号密码
6. SQLServer2008快速清理日志文件
7. switch 根据键盘录入成绩 显示分数及判断等级（第二次）
8. 内网渗透-域内有网和无网
9. [网络安全自学篇] 十五.Python攻防之多线程、C段扫描和数据库编程（二）
10. 【数据结构与算法】之深入解析“序列化和反序列化二叉搜索树”的求解思路与算法示例