Quartus Embedding Layer #548
Conversation
hls4ml/templates/quartus/firmware/nnet_utils/nnet_embed_stream.h
Outdated
Show resolved
Hide resolved
| #ifndef NNET_BATCHNORM_STREAM_H_ | ||
| #define NNET_BATCHNORM_STREAM_H_ | ||
|
|
||
| namespace nnet {} |
There was a problem hiding this comment.
Can we add a stub that raises an exception while we still don't have the implementation? Otherwise the compilation fails with a cryptic message (cannot substitute template argument or something like that). Or does that require the more streaming infrastructure in Quartus (#557)?
There was a problem hiding this comment.
This file should never be invoked, as soon as Quartus attempts to use io_stream an exception is thrown from Python (although #557 attempts to add infrastructure for streaming, which means for some layers we should throw exceptions and for some not - this should be handled by Python and can be a part of the streaming PR)
| DenseEmbedding: | ||
| #pragma unroll | ||
| for (int i = 0; i < CONFIG_T::n_out; i++) { | ||
| res[j * CONFIG_T::n_out + i] = embeddings[data[j].to_uint() * CONFIG_T::n_out + i]; |
There was a problem hiding this comment.
Remind me, why is the to_uint() needed here?
There was a problem hiding this comment.
data[j] is a fixed point number by default, without to_uint() compilation will fail as ac_fixed and int are not interchangeable - may be allowed with the -fpermissive flag, not sure.
Added support for
Embbedinglayer in Quartus, similar to work in this commit: 5040f75#diff-7dd4e5f5df1fd1a2cf2974ce0982818da375ff46be0eda702ba60b4f6188d928Will be useful once work on
SimpleRNNandLSTMis completed.