普通卷积armv7-neon指令集实现

4x8的分块，很简单，主要看一下所用的指令集


void xnet_f32_igemm_ukernel_4x8__neon_lane_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const float** a,const float* w,float* c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float* zero)
{assert(mr != 0);assert(mr <= 4);assert(nc != 0);assert(kc != 0);assert(kc % sizeof(float) == 0);assert(ks != 0);assert(ks % (4 * sizeof(void*)) == 0);assert(a_offset % sizeof(float) == 0);assert(a != NULL);assert(w != NULL);assert(c != NULL);float* c0 = c;float* c1 = (float*) ((uintptr_t) c0 + cm_stride);if XNET_UNPREDICTABLE(mr < 2) {c1 = c0;}float* c2 = (float*) ((uintptr_t) c1 + cm_stride);if XNET_UNPREDICTABLE(mr <= 2) {c2 = c1;}float* c3 = (float*) ((uintptr_t) c2 + cm_stride);if XNET_UNPREDICTABLE(mr != 4) {c3 = c2;}do {float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;float32x4_t vacc1x0123 = vacc0x0123;float32x4_t vacc1x4567 = vacc0x4567;float32x4_t vacc2x0123 = vacc0x0123;float32x4_t vacc2x4567 = vacc0x4567;float32x4_t vacc3x0123 = vacc0x0123;float32x4_t vacc3x4567 = vacc0x4567;size_t p = ks;do {const float* a0 = a[0];assert(a0 != NULL);if XNET_UNPREDICTABLE(a0 != zero) {a0 = (const float*) ((uintptr_t) a0 + a_offset);}const float* a1 = a[1];assert(a1 != NULL);if XNET_UNPREDICTABLE(a1 != zero) {a1 = (const float*) ((uintptr_t) a1 + a_offset);}const float* a2 = a[2];assert(a2 != NULL);if XNET_UNPREDICTABLE(a2 != zero) {a2 = (const float*) ((uintptr_t) a2 + a_offset);}const float* a3 = a[3];assert(a3 != NULL);if XNET_UNPREDICTABLE(a3 != zero) {a3 = (const float*) ((uintptr_t) a3 + a_offset);}a += 4;size_t k = kc;for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {const float32x4_t va0 = vld1q_f32(a0); a0 += 4;const float32x4_t va1 = vld1q_f32(a1); a1 += 4;const float32x4_t va2 = vld1q_f32(a2); a2 += 4;const float32x4_t va3 = vld1q_f32(a3); a3 += 4;const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, vget_low_f32(va0), 0);vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, vget_low_f32(va1), 0);vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, vget_low_f32(va2), 0);vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, vget_low_f32(va3), 0);vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, vget_low_f32(va0), 0);vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, vget_low_f32(va1), 0);vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, vget_low_f32(va2), 0);vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, vget_low_f32(va3), 0);const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, vget_low_f32(va0), 1);vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, vget_low_f32(va1), 1);vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, vget_low_f32(va2), 1);vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, vget_low_f32(va3), 1);vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c1, vget_low_f32(va0), 1);vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c1, vget_low_f32(va1), 1);vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c1, vget_low_f32(va2), 1);vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c1, vget_low_f32(va3), 1);const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c2, vget_high_f32(va0), 0);vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c2, vget_high_f32(va1), 0);vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c2, vget_high_f32(va2), 0);vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c2, vget_high_f32(va3), 0);vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c2, vget_high_f32(va0), 0);vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c2, vget_high_f32(va1), 0);vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c2, vget_high_f32(va2), 0);vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c2, vget_high_f32(va3), 0);const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c3, vget_high_f32(va0), 1);vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c3, vget_high_f32(va1), 1);vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c3, vget_high_f32(va2), 1);vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c3, vget_high_f32(va3), 1);vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c3, vget_high_f32(va0), 1);vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c3, vget_high_f32(va1), 1);vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c3, vget_high_f32(va2), 1);vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c3, vget_high_f32(va3), 1);}if XNET_UNLIKELY(k != 0) {do {const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;const float32x4_t vb0123 = vld1q_f32(w); w += 4;const float32x4_t vb4567 = vld1q_f32(w); w += 4;vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);k -= sizeof(float);} while (k != 0);}p -= 4 * sizeof(void*);} while (p != 0);if XNET_LIKELY(nc >= 8) {vst1q_f32(c3, vacc3x0123);vst1q_f32(c3 + 4, vacc3x4567);c3 = (float*) ((uintptr_t) c3 + cn_stride);vst1q_f32(c2, vacc2x0123);vst1q_f32(c2 + 4, vacc2x4567);c2 = (float*) ((uintptr_t) c2 + cn_stride);vst1q_f32(c1, vacc1x0123);vst1q_f32(c1 + 4, vacc1x4567);c1 = (float*) ((uintptr_t) c1 + cn_stride);vst1q_f32(c0, vacc0x0123);vst1q_f32(c0 + 4, vacc0x4567);c0 = (float*) ((uintptr_t) c0 + cn_stride);a = (const float**) ((uintptr_t) a - ks);nc -= 8;} else {if (nc & 4) {vst1q_f32(c3, vacc3x0123); c3 += 4;vst1q_f32(c2, vacc2x0123); c2 += 4;vst1q_f32(c1, vacc1x0123); c1 += 4;vst1q_f32(c0, vacc0x0123); c0 += 4;vacc3x0123 = vacc3x4567;vacc2x0123 = vacc2x4567;vacc1x0123 = vacc1x4567;vacc0x0123 = vacc0x4567;}float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);if (nc & 2) {vst1_f32(c3, vacc3x01); c3 += 2;vst1_f32(c2, vacc2x01); c2 += 2;vst1_f32(c1, vacc1x01); c1 += 2;vst1_f32(c0, vacc0x01); c0 += 2;vacc3x01 = vget_high_f32(vacc3x0123);vacc2x01 = vget_high_f32(vacc2x0123);vacc1x01 = vget_high_f32(vacc1x0123);vacc0x01 = vget_high_f32(vacc0x0123);}if (nc & 1) {vst1_lane_f32(c3, vacc3x01, 0);vst1_lane_f32(c2, vacc2x01, 0);vst1_lane_f32(c1, vacc1x01, 0);vst1_lane_f32(c0, vacc0x01, 0);}nc = 0;}} while (nc != 0);
}

所用指令集：

vld1q_f32(const float*)
vmlaq_lane_f32(float32x4_t a, float32x4_t b, float32x2 c, int lane): a=a+b*c[lane]
float32x2_t vget_low_f32/vget_high_f32(float32x4_t)
vld1q_dup_f32(const float*)
vmlaq_f32(float32x4_t a, float32x4_t b, float32x4 c): a=a+b*c
vst1q_f32(const float* , float32x4_t)
vst1_f32(const float* , float32x2_t)
vst1_lane_f32(const float*, float32x2_t, int lane)

普通卷积armv7-neon指令集实现—XNNPACK相关推荐

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 ...
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 ...
【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. ※ ...
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 ...
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 ...
【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 ...
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 ...
【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 ...
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 ...

普通卷积armv7-neon指令集实现—XNNPACK

普通卷积armv7-neon指令集实现—XNNPACK相关推荐

最新文章

热门文章