Description

NDK r23b compile neon intrinsics is very slow:

https://github.com/Tencent/ncnn/blob/master/src/mat_pixel_affine.cpp#L1173

warpaffine_bilinear_c4

params:
constexpr int width = 160;
constexpr int height = 160;
constexpr float image_matrix[6] = {
-0.00673565036, 0.146258384, 4.34562492,
-0.146258384, -0.00673565036, 162.753372,
};

NDK r23b this funciton cost 8.40 ms.
NDK r22b this function cost 0.302 ms.

Environment Details

• NDK Version: r23b
• Build system: CMake
• Host OS: Mac
• ABI: arm64-v8a
• NDK API level: android-21
• Device API level: android-30