Skip to content

Commit

Permalink
[ntuple] add paragraph to architecture on low-prec floats
Browse files Browse the repository at this point in the history
  • Loading branch information
silverweed committed Sep 24, 2024
1 parent 984bc4a commit d6c9c0f
Show file tree
Hide file tree
Showing 2 changed files with 46 additions and 5 deletions.
41 changes: 41 additions & 0 deletions tree/ntuple/v7/doc/architecture.md
Original file line number Diff line number Diff line change
Expand Up @@ -450,6 +450,47 @@ Every fill context prepares a set of entire clusters in the final on-disk layout
When a fill context flushes data,
a brief serialization point handles the RNTuple meta-data updates and the reservation of disk space to write into.
Low precision float types
--------------------------
RNTuple supports encoding floating point types with a lower precision when writing them to disk. This encoding is specified by the
user per field and it is independent on the in-memory type used for that field (meaning both a `RField<double>` or `RField<float>` can
be mapped to e.g. a low-precision 16 bit float).
RNTuple supports the following encodings (all mutually exclusive):
- **Real16**/**SplitReal16**: IEEE-754 half precision float. Set by calling `RField::SetHalfPrecision()`;
- **Real32Trunc**: floating point with less than 32 bits of precision (truncated mantissa).
Set by calling `RField::SetTruncated(n)`, with $10 <= n <= 31$ equal to the total number of bits used on disk.
Note that `SetTruncated(16)` makes this effectively a `bfloat16` on disk;
- **Real32Quant**: floating point with a normalized/quantized integer representation on disk using a user-specified number of bits.
Set by calling `RField::SetQuantized(min, max, nBits)`, where $1 <= nBits <= 32$.
This representation will map the floating point value `min` to 0, `max` to the highest representable integer with `nBits` and any
value in between will be a linear interpolation of the two. It is up to the user to ensure that only values between `min` and `max`
are stored in this field. The current RNTuple implementation will throw an exception if that is not the case when writing the values to disk.
In addition to these encodings, a user may call `RField<double>::SetDouble32()` to set the column representation of a `double` field to
a 32-bit floating point value. The default behavior of `Float16_t` can be emulated by calling `RField::SetTruncated(21)` (which will truncate
a single precision float's mantissa to 12 bits).
Here is an example on how a user may dynamically decide how to quantize a floating point field to get the most precision out of a fixed bit width:
```c++
auto model = RNTupleModel::Create();
auto field = std::make_unique<RField<float>>("f");
// assuming we have an array of floats stored in `myFloats`:
auto [minV, maxV] = std::minmax_element(myFloats.begin(), myFloats.end());
constexpr auto nBits = 24;
field->SetQuantized(*minV, *maxV, nBits);
model->AddField(std::move(field));
auto f = model->GetDefaultEntry().GetPtr<float>("f");
// Now we can write our floats.
auto writer = RNTupleWriter::Recreate(std::move(model), "myNtuple", "myFile.root");
for (float val : myFloats) {
*f = val;
writer->Fill();
}
```

Relationship to other ROOT components
-------------------------------------

Expand Down
10 changes: 5 additions & 5 deletions tree/ntuple/v7/src/RColumnElement.hxx
Original file line number Diff line number Diff line change
Expand Up @@ -788,12 +788,12 @@ int QuantizeReals(Quantized_t *dst, const T *src, std::size_t count, double min,
int nOutOfRange = 0;

for (std::size_t i = 0; i < count; ++i) {
T elem = src[i];
const T elem = src[i];

nOutOfRange += !(min <= elem && elem <= max);

double e = (elem - min) * scale;
Quantized_t q = static_cast<Quantized_t>(e + 0.5);
const double e = 0.5 + (elem - min) * scale;
Quantized_t q = static_cast<Quantized_t>(e);
ByteSwapIfNecessary(q);

// double-check we actually used at most `nQuantBits`
Expand Down Expand Up @@ -830,8 +830,8 @@ int UnquantizeReals(T *dst, const Quantized_t *src, std::size_t count, double mi
elem >>= unusedBits;
ByteSwapIfNecessary(elem);

double fq = static_cast<double>(elem);
double e = (fq + bias) * scale;
const double fq = static_cast<double>(elem);
const double e = (fq + bias) * scale;
dst[i] = static_cast<T>(e);

nOutOfRange += !(min <= dst[i] && dst[i] <= max);
Expand Down

0 comments on commit d6c9c0f

Please sign in to comment.