| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| |
|
| | #include "clip_mips.h" |
| |
|
| | #if __mips_msa |
| | #include <msa.h> |
| | #endif |
| |
|
| | #include "mips_usability.h" |
| |
|
| | namespace ncnn { |
| |
|
| | Clip_mips::Clip_mips() |
| | { |
| | #if __mips_msa |
| | support_packing = true; |
| | #endif |
| | } |
| |
|
| | int Clip_mips::forward_inplace(Mat& bottom_top_blob, const Option& opt) const |
| | { |
| | int w = bottom_top_blob.w; |
| | int h = bottom_top_blob.h; |
| | int d = bottom_top_blob.d; |
| | int channels = bottom_top_blob.c; |
| | int elempack = bottom_top_blob.elempack; |
| | int size = w * h * d * elempack; |
| |
|
| | #pragma omp parallel for num_threads(opt.num_threads) |
| | for (int q = 0; q < channels; q++) |
| | { |
| | float* ptr = bottom_top_blob.channel(q); |
| |
|
| | int i = 0; |
| | #if __mips_msa |
| | v4f32 _max = (v4f32)__msa_fill_w_f32(max); |
| | v4f32 _min = (v4f32)__msa_fill_w_f32(min); |
| | for (; i + 3 < size; i += 4) |
| | { |
| | __builtin_prefetch(ptr + 16); |
| | v4f32 _p = (v4f32)__msa_ld_w(ptr, 0); |
| | _p = __msa_fmax_w(_p, _min); |
| | _p = __msa_fmin_w(_p, _max); |
| | __msa_st_w((v4i32)_p, ptr, 0); |
| |
|
| | ptr += 4; |
| | } |
| | #endif |
| | for (; i < size; i++) |
| | { |
| | if (*ptr < min) |
| | *ptr = min; |
| |
|
| | if (*ptr > max) |
| | *ptr = max; |
| |
|
| | ptr++; |
| | } |
| | } |
| |
|
| | return 0; |
| | } |
| |
|
| | } |
| |
|