Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2297
Star 20931
Fork 5422
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 c54c7d91 编写于 作者: L Liu Yiqun
浏览文件
操作

Use template to deliver const argument instead, to remove the compiling error... 

Use template to deliver const argument instead, to remove the compiling error "argument to __builtin_neon_vgetq_lane_f32 must be a constant integer".
上级 50c63dc9
隐藏空白更改
内联 并排
Showing with 52 addition and 52 deletion
paddle/function/neon/NeonDepthwiseConv.cpp
浏览文件 @ c54c7d91
......@@ -116,15 +116,15 @@ struct DepthwiseConvKernel<3, 1> {
float32x4_t tmp1 = vdupq_n_f32(0.f);
float32x4_t tmp2 = vdupq_n_f32(0.f);
tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0);
tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1);
tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2);
tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[0][0], k[0]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[0][1], k[0]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[0][2], k[0]);
tmp2 = vmlaq_laneq_f32<0>(tmp2, input[1][0], k[1]);
tmp1 = vmlaq_laneq_f32<1>(tmp1, input[1][1], k[1]);
tmp2 = vmlaq_laneq_f32<2>(tmp2, input[1][2], k[1]);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[2][0], k[2]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[2][1], k[2]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[2][2], k[2]);
tmp1 = vaddq_f32(tmp1, tmp2);
vst1q_f32(outputData, tmp1);
......@@ -223,15 +223,15 @@ struct DepthwiseConvKernel<3, 2> {
float32x4_t tmp1 = vdupq_n_f32(0.f);
float32x4_t tmp2 = vdupq_n_f32(0.f);
tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0);
tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1);
tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2);
tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[0][0], k[0]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[0][1], k[0]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[0][2], k[0]);
tmp2 = vmlaq_laneq_f32<0>(tmp2, input[1][0], k[1]);
tmp1 = vmlaq_laneq_f32<1>(tmp1, input[1][1], k[1]);
tmp2 = vmlaq_laneq_f32<2>(tmp2, input[1][2], k[1]);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[2][0], k[2]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[2][1], k[2]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[2][2], k[2]);
tmp1 = vaddq_f32(tmp1, tmp2);
vst1q_f32(outputData, tmp1);
......@@ -316,22 +316,22 @@ struct DepthwiseConvKernel<4, 1> {
float32x4_t tmp1 = vdupq_n_f32(0.f);
float32x4_t tmp2 = vdupq_n_f32(0.f);
tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3);
tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3);
tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3);
tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[0][0], k[0]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[0][1], k[0]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[0][2], k[0]);
tmp2 = vmlaq_laneq_f32<3>(tmp2, input[0][3], k[0]);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[1][0], k[1]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[1][1], k[1]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[1][2], k[1]);
tmp2 = vmlaq_laneq_f32<3>(tmp2, input[1][3], k[1]);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[2][0], k[2]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[2][1], k[2]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[2][2], k[2]);
tmp2 = vmlaq_laneq_f32<3>(tmp2, input[2][3], k[2]);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[3][0], k[3]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[3][1], k[3]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[3][2], k[3]);
tmp2 = vmlaq_laneq_f32<3>(tmp2, input[3][3], k[3]);
tmp1 = vaddq_f32(tmp1, tmp2);
vst1q_f32(outputData, tmp1);
......@@ -431,22 +431,22 @@ struct DepthwiseConvKernel<4, 2> {
float32x4_t tmp1 = vdupq_n_f32(0.f);
float32x4_t tmp2 = vdupq_n_f32(0.f);
tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3);
tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3);
tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3);
tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0);
tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1);
tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2);
tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[0][0], k[0]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[0][1], k[0]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[0][2], k[0]);
tmp2 = vmlaq_laneq_f32<3>(tmp2, input[0][3], k[0]);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[1][0], k[1]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[1][1], k[1]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[1][2], k[1]);
tmp2 = vmlaq_laneq_f32<3>(tmp2, input[1][3], k[1]);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[2][0], k[2]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[2][1], k[2]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[2][2], k[2]);
tmp2 = vmlaq_laneq_f32<3>(tmp2, input[2][3], k[2]);
tmp1 = vmlaq_laneq_f32<0>(tmp1, input[3][0], k[3]);
tmp2 = vmlaq_laneq_f32<1>(tmp2, input[3][1], k[3]);
tmp1 = vmlaq_laneq_f32<2>(tmp1, input[3][2], k[3]);
tmp2 = vmlaq_laneq_f32<3>(tmp2, input[3][3], k[3]);
tmp1 = vaddq_f32(tmp1, tmp2);
vst1q_f32(outputData, tmp1);
......
paddle/function/neon/neon_util.h
浏览文件 @ c54c7d91
......@@ -33,10 +33,10 @@ inline float32_t vaddvq_f32(float32x4_t a) {
return vget_lane_f32(vpadd_f32(v, v), 0);
}
template <int lane>
inline float32x4_t vmlaq_laneq_f32(float32x4_t a,
float32x4_t b,
float32x4_t v,
const int lane) {
float32x4_t v) {
return vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane));
}
#endif
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或