public static void Activation_Sigmoid(Tensor input, Tensor output)
{
__m256 one = Vector256.Create(1.0f);
for (int i = 0; i < output.xmm.Length; ++i)
{
output.xmm[i] = NN_utils.exp256_ps(input.xmm[i]);
var divisor = Avx2.Add(output.xmm[i], one);
output.xmm[i] = Avx2.Divide(output.xmm[i], divisor);
}
}
public static void Activation_Softmax(Tensor input, Tensor output)
{
float sum = 0.0f;
for (int i = 0; i < output.xmm.Length; ++i)
{
output.xmm[i] = NN_utils.exp256_ps(input.xmm[i]);
sum += NN_utils.hsums(output.xmm[i]).GetElement <float>(0);
}
sum = 1.0f / sum;
output.mul(sum, output);
}
public void dot_product(int kept_dim, List <float> big_matrix_vec, int big_reserve_size,
Tensor small, int chunk_range, Tensor o)
{
int out_index = 0;
for (int small_chunk = 0; small_chunk < small.xmm.Length; small_chunk += chunk_range)
{
for (int big_chunk = 0; big_chunk < xmm.Length; big_chunk += chunk_range)
{
__m256 FMA = Vector256 <float> .Zero;
for (int partial_index = 0; partial_index < chunk_range; ++partial_index)
{
FMA = Fma.MultiplyAdd(xmm[big_chunk + partial_index], small.xmm[small_chunk + partial_index], FMA);
}
o.setElement(out_index++, NN_utils.hsums(FMA).GetElement <float>(0));
}
}
}
