add Force() to force computation of histogram if not cached

[tbonfort]

No description provided.

[coveralls]

Pull Request Test Coverage Report for Build 1602495501

• 7 of 7 (100.0%) changed or added relevant lines in 2 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.0008%) to 99.648%

---

Totals
Change from base Build 1520238762:	0.0008%
Covered Lines:	2550
Relevant Lines:	2559

---

💛 - Coveralls

[tbonfort]

cc @fhelen for the addition of Force() to be reused by Statistics()

[tbonfort]

cc @thomascoquet : this should fix the issue you raised in #60 about Histogram() forcing a full scan.

[tbonfort]

@thomascoquet using Histogram(Intervals(count,min,max)) will always force a dataset scan. Should we require a user to write Histogram(Intervals(count,min,max), Force()), i.e. return an error if Force is not provided?

[thomascoquet]

Sorry @tbonfort, I missed the GitHub notification in my mails. I agree it is better to have an adequate flag such as Force() ; last thing one wants is a unintentional remote read.

---

I think it is a little bit obscure how to properly use histograms in godal. Indeed, there are two ways to compute histograms:

• one uses band->GetDefaultHistogram that returns a cached version or recompute using full raster and never uses Approximate() (c.f. https://github.com/OSGeo/gdal/blob/master/gcore/gdalrasterband.cpp#L3532)
• another uses band->GetHistogram that is capable of using approximation but does not look for cached histograms.

A very (and maybe the most) frequent use case for histograms is what I would call the "QGIS" use case: there is a raster you know nothing about, you need to display some color palette on it and apply some percentile-based stretching. In which case you would like the cached histogram or a cheap approximated version.

The most natural way is to call Histogram() and maybe Histogram(Approx(), Force()) to allow computation on overviews if not cached. With the current flow, Histogram(Approx(), Force()) will never consider the Approx option and will read the raster at full resolution.

Therefore, I would return an error in this PR if both options are used and wait for @fhelen to solve this using approximated statistics.