Sha256: 0e3bf5d2fc179875c4b23807cb1a5eba1dbde8245dac176ba404fe5b2f6a7fed
Contents?: true
Size: 797 Bytes
Versions: 20
Compression:
Stored size: 797 Bytes
Contents
// First naive implementation
% c_dtype = dtype_to_c_type(dtype)
__kernel void log_softmax_<%= dtype %>(const int N,
const __global <%= c_dtype %>* A,
__global <%= c_dtype %>* C) {
// Get the index of the current element to be processed
const int globalRow = get_global_id(0); // Row ID of C (0..M)
// Compute a single element (loop over K)
<%= c_dtype %> acc = 0.0f;
<%= c_dtype %> max = <%= min_value_for(dtype) %>;
for (int k=0; k<N; k++) {
max = A[globalRow*N + k] > max ? A[globalRow*N + k] : max;
}
for (int k=0; k<N; k++) {
acc += exp(A[globalRow*N + k] - max);
}
// Store the result
for (int k=0; k < N; k++) {
C[globalRow*N + k] = (A[globalRow*N + k] - max) - log(acc);
}
}
Version data entries
20 entries across 20 versions & 2 rubygems