Reimplement managed raster class in C++ #1597

emlys · 2024-07-17T17:59:19Z

Description

This PR reimplements the ManagedRaster class in C++ for better performance. Tested with sdr.calculate_sediment_deposition and the SDR sample data, it's now just slightly faster than the original implementation on main (0.59 vs 0.61 seconds), whereas the python implementation on feature/routing-refactor takes nearly twice as long.

I'm a beginner at C++ so I'm sure I'm not always using the right conventions or design patterns, but this seems like an okay starting point. I took some guidance from the Google C++ style guide. Ignoring some guidelines that don't seem to make sense in this context. For instance, I have all the code in a header file because splitting it out into ManagedRaster.h and ManagedRaster.cpp significantly slowed it down.

Happy to discuss and walk through this PR on a zoom call as well.

Checklist

Updated HISTORY.rst and link to any relevant issue (if these changes are user-facing)
Updated the user's guide (if needed)
Tested the Workbench UI (if relevant)

…-compare

phargogh

This is looking really great @emlys ! I had a few questions, but I think they're all pretty minor. Happy to talk through anything that would be helpful!

phargogh · 2024-09-03T20:10:48Z

setup.py

 if platform.system() == 'Darwin':
-    compiler_and_linker_args = ['-stdlib=libc++']
+    compiler_args = []
+    compiler_and_linker_args = ['-stdlib=libc++', '-std=c++20']
+    library_dirs = [f'{os.environ["CONDA_PREFIX"]}/lib']
+else:
+    compiler_args = ['/std:c++20']
+    library_dirs = [f'{os.environ["CONDA_PREFIX"]}/Library/lib']
+    include_dirs.append(f'{os.environ["CONDA_PREFIX"]}/Library/include')


Since it's possible that someone would compile on Linux, would the /srd:c++20 flag for with GNU? Or do we need to handle the linux possibility separately?

Have we been supporting Linux compilation previously? I think this would not work, but if it didn't work before, maybe that's a separate question.

Yeah, it's never been a priority in this repository, but we do offer linux builds via conda-forge so I don't think we should completely ignore the use case. Linux support has historically been kind of informal in this repo.

phargogh · 2024-09-03T21:11:57Z

src/natcap/invest/managed_raster/ManagedRaster.h

+            raster_x_size = dataset->GetRasterXSize();
+            raster_y_size = dataset->GetRasterYSize();
+
+            if (band_id < 1 or band_id > dataset->GetRasterCount()) {


Fascinating, I wasn't aware that or was a keyword here in C++! Looks like their alternative keywords can be pretty cryptic, but this one is quite readable.

phargogh · 2024-09-03T21:32:46Z

src/natcap/invest/managed_raster/ManagedRaster.h

+            closed = 0;
+        }
+
+        void set(long xi, long yi, double value) {


I think I remember Rich using the inline keyword here as well to try and eke out some minor speedup since the function was so small. Is there any noticeable speedup here by using void inline set(...? There might not be, and the function is long enough where there probably isn't a noticeable speedup.

I'm not seeing a measurable speedup on the SDR sample data from inlining this (or the other function below). As a guideline, the Google C++ style guide recommends against inlining functions over about 10 lines, so these are on the high end of that. But since they were originally inlined in the Cython version, I don't see any harm in making them inlined here!

-> Caveat that I'm not at all confident that my speed testing will detect relatively small changes. I get a lot of variation in times.

src/natcap/invest/managed_raster/ManagedRaster.h

phargogh · 2024-09-03T21:33:47Z

src/natcap/invest/managed_raster/ManagedRaster.h

+            }
+        }
+
+        double get(long xi, long yi) {


Same question here about whether there's any speedup from adding the inline keyword.

See comment above

phargogh · 2024-09-03T23:04:46Z

src/natcap/invest/ndr/ndr_core.pyx

+    cdef ManagedRaster stream_raster = ManagedRaster(
+        stream_path.encode('utf-8'), 1, False)


I imagine once we have the python C interface for logging figured out that we might be able to do this string encoding as well!

phargogh · 2024-09-03T23:13:39Z

src/natcap/invest/ndr/ndr_core.pyx

+            neighbor = up_iterator.next()
+            while neighbor.direction < 8:


I remember one of the things we talked about last week was the need to be careful to cover all of the iteration cases (e.g. continue, end of loop block) with calling up_iterator.next(). What if we just handled that up here at the top of the loop? I suppose there's a slight cost to holding 2 neighbor objects in memory, but maybe it would be easier to maintain? I'm thinking of something like the below (github wouldn't allow me to suggest something on the whole loop block)

Suggested change

neighbor = up_iterator.next()

while neighbor.direction < 8:

neighbor = up_iterator.next()

next_neighbor = neighbor

while neighbor.direction < 8:

neighbor = next_neighbor

next_neighbor = up_iterator.next()

@phargogh I think I've found an even better solution which is to properly implement the iterators to meet the C++ iterator specification. That way we don't need to manually call next() at all. I was stumped on that before but I think I've worked it out.

Woohoo! That is super exciting. Well done and I can't wait to see the iterator!

phargogh · 2024-09-03T23:15:10Z

src/natcap/invest/sdr/sdr_core.pyx

+                    up_iterator = UpslopeNeighborIterator(
+                        mfd_flow_direction_raster, global_col, global_row)
+                    neighbor = up_iterator.next()
+                    while neighbor.direction < 8:


Same question here about whether it'd make sense to track a neighbor/next neighbor object in the loop to make it easier on us to maintain and avoid infinite loops.

phargogh · 2024-09-03T23:17:33Z

src/natcap/invest/sdr/sdr_core_main.pyx

I remember we talked about this file, but I forget if it's intended to continue to be in the PR. Should this file be removed as a part of the PR or should it be retained for now?

Oops, it's meant to be removed!

phargogh · 2024-09-03T23:19:11Z

src/natcap/invest/seasonal_water_yield/seasonal_water_yield_core.pyx

+    for i, et0_path in enumerate(et0_path_list):
+        et0_m_rasters.push_back(ManagedRaster(et0_path.encode('utf-8'), 1, 0))


Is the enumeration still needed? It looks like there are a few more enumerations in the next 30 lines or so

Not needed!

Co-authored-by: James Douglass <[email protected]>

…-compare

emlys added 5 commits May 7, 2024 10:25

compare different implementations of upslope neighbors

5e58f60

improve efficiency of managedraster helper functions

7c51c08

work in progress: rewrite managed raster module in c++

a55d57c

work in progress: managed raster in C++

068d886

C++ managed raster implementation now faster on SDR sample data

8576088

emlys changed the base branch from main to feature/routing-refactor July 17, 2024 18:00

emlys added 24 commits July 17, 2024 11:18

Merge branch 'feature/routing-refactor' into feature/routing-refactor…

7f2ec80

…-compare

NDR tests passing with new C++ implementation of managedraster

11ef40f

all tests using managedraster passing locally

c7486f2

debugging: find gdal.h

048e769

debugging: add incldue path to compile args for windows

cccf0e6

debugging

8d104a4

debugging'

6ca0f78

debugging

61d1f65

debugging

d3aebac

debugging

eb15fa1

debugging

93ec4b9

debugging

4506414

debugging

1ff72b6

debugging

6acaea6

use C++20; no need to compile managed_raster as a separate extension

9316d20

debugging

0d4805a

debugging

53b15d0

debugging

06c8648

debugging

1749eb9

debugging

e4c234f

debugging

bbd7c5a

debugging

7bc4b58

debugging

efc5527

debugging

474a89a

phargogh reviewed Sep 3, 2024

View reviewed changes

emlys and others added 29 commits October 15, 2024 13:42

Update src/natcap/invest/managed_raster/ManagedRaster.h

364d7ff

Co-authored-by: James Douglass <[email protected]>

remove unneeded file

386beb2

remove unused enumeration

7b994e2

inline managedraster set and get functions

f62a696

write C++ error messages out to error stream

8109a60

clean up lru_cache and actualBlockWidths in managedraster.close

da23f3e

inline is_close function

31565e8

use gdal-config to get library path on macos

8c31a12

separate C++ classes for each variant of neighbor pixel iteration

b88567c

drafting a true C++ iterator

36c26b2

implement true C++ iterator for downslope neighbors and use in SDR

fc8dd38

using downslope iterators everywhere

e6a370e

use true iterators for upslope neighbors

2cf8f5d

Merge branch 'feature/routing-refactor' into feature/routing-refactor…

ba37a0a

…-compare

debugging

b957881

debugging

989a7b6

cdef flow_dir_sum

b8e710f

debugging

b532001

debugging

9421e2c

debugging

9981569

debugging

6d556e2

debugging

c7742e8

debugging

7b9db1f

debugging

9071362

debugging

ba9cebe

debugging

13cc0dc

use const reference parameters to equality fns

7a767f5

inherit from Neighbors class

5b8c04e

inherit operator funcs

947900c

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Reimplement managed raster class in C++ #1597

Reimplement managed raster class in C++ #1597

emlys commented Jul 17, 2024 •

edited

Loading

phargogh left a comment

phargogh Sep 3, 2024

emlys Oct 23, 2024

phargogh Oct 23, 2024

phargogh Sep 3, 2024

phargogh Sep 3, 2024

emlys Oct 15, 2024 •

edited

Loading

phargogh Sep 3, 2024

emlys Oct 15, 2024

phargogh Sep 3, 2024

phargogh Sep 3, 2024

emlys Oct 29, 2024

phargogh Oct 29, 2024

phargogh Sep 3, 2024

phargogh Sep 3, 2024

emlys Oct 15, 2024

phargogh Sep 3, 2024

emlys Oct 15, 2024

		cdef ManagedRaster stream_raster = ManagedRaster(
		stream_path.encode('utf-8'), 1, False)

		for i, et0_path in enumerate(et0_path_list):
		et0_m_rasters.push_back(ManagedRaster(et0_path.encode('utf-8'), 1, 0))

Reimplement managed raster class in C++ #1597

Are you sure you want to change the base?

Reimplement managed raster class in C++ #1597

Conversation

emlys commented Jul 17, 2024 • edited Loading

Description

Checklist

phargogh left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

emlys Oct 15, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

emlys commented Jul 17, 2024 •

edited

Loading

emlys Oct 15, 2024 •

edited

Loading