| 24 | namespace ipl { |
| 25 | |
| 26 | void ArrayMath::add(int size, |
| 27 | const float* in1, |
| 28 | const float* in2, |
| 29 | float* out) |
| 30 | { |
| 31 | auto simdSize = size & ~3; |
| 32 | |
| 33 | if (float4::isAligned(in1) && float4::isAligned(in2) && float4::isAligned(out)) |
| 34 | { |
| 35 | for (auto i = 0; i < simdSize; i += 4) |
| 36 | { |
| 37 | float4::store(&out[i], float4::add(float4::load(&in1[i]), float4::load(&in2[i]))); |
| 38 | } |
| 39 | } |
| 40 | else |
| 41 | { |
| 42 | for (auto i = 0; i < simdSize; i += 4) |
| 43 | { |
| 44 | float4::storeu(&out[i], float4::add(float4::loadu(&in1[i]), float4::loadu(&in2[i]))); |
| 45 | } |
| 46 | } |
| 47 | |
| 48 | for (auto i = simdSize; i < size; ++i) |
| 49 | { |
| 50 | out[i] = in1[i] + in2[i]; |
| 51 | } |
| 52 | } |
| 53 | |
| 54 | void ArrayMath::add(int size, |
| 55 | const complex_t* in1, |