| 148 | |
| 149 | template <typename T, class ExpFunc> |
| 150 | void pow_invoke(const T* src, T* dst, size_t size, ExpFunc expfunc) { |
| 151 | size_t i; |
| 152 | for (i = 0; i + 4 <= size; i += 4) { |
| 153 | T a0 = src[i], a1 = src[i + 1], a2 = src[i + 2], a3 = src[i + 3]; |
| 154 | T b0 = expfunc.apply(a0), b1 = expfunc.apply(a1), b2 = expfunc.apply(a2), |
| 155 | b3 = expfunc.apply(a3); |
| 156 | dst[i] = b0; |
| 157 | dst[i + 1] = b1; |
| 158 | dst[i + 2] = b2; |
| 159 | dst[i + 3] = b3; |
| 160 | } |
| 161 | #if MEGDNN_FIX_AARCH32_BUG |
| 162 | // FIXME: as llvm may cause cannot select error if enable vectorize |
| 163 | #pragma clang loop vectorize(disable) |
| 164 | #endif |
| 165 | for (; i < size; ++i) { |
| 166 | dst[i] = expfunc.apply(src[i]); |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | bool float_eq(float x, float y) { |
| 171 | return std::abs(x - y) < std::numeric_limits<float>::epsilon(); |