compiler: Improve data streaming backend #2390

FabioLuporini · 2024-06-19T12:46:02Z

Yet another bunch of misc compiler tweaks and generalisations, mostly to better support lazy data streaming in PRO

devito/ir/stree/algorithms.py

codecov · 2024-06-19T12:49:18Z

Codecov Report

Attention: Patch coverage is 52.97619% with 79 lines in your changes missing coverage. Please review.

Project coverage is 86.69%. Comparing base (e37d6ff) to head (4d08839).

Files	Patch %	Lines
devito/operator/operator.py	43.95%	49 Missing and 2 partials ⚠️
tests/test_gpu_common.py	0.00%	7 Missing ⚠️
devito/passes/iet/instrument.py	20.00%	4 Missing ⚠️
devito/types/basic.py	66.66%	4 Missing ⚠️
devito/types/parallel.py	66.66%	3 Missing and 1 partial ⚠️
devito/mpi/halo_scheme.py	40.00%	3 Missing ⚠️
devito/ir/clusters/algorithms.py	66.66%	1 Missing and 1 partial ⚠️
devito/ir/stree/algorithms.py	0.00%	2 Missing ⚠️
devito/arch/archinfo.py	66.66%	1 Missing ⚠️
devito/types/dense.py	75.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2390      +/-   ##
==========================================
- Coverage   86.78%   86.69%   -0.09%     
==========================================
  Files         234      234              
  Lines       44271    44383     +112     
  Branches     8184     8216      +32     
==========================================
+ Hits        38419    38477      +58     
- Misses       5137     5185      +48     
- Partials      715      721       +6

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

georgebisbas · 2024-06-19T14:43:14Z

devito/arch/archinfo.py

@@ -19,6 +19,8 @@
           'get_cuda_path', 'get_hip_path', 'check_cuda_runtime', 'get_m1_llvm_path',
           'Platform', 'Cpu64', 'Intel64', 'IntelSkylake', 'Amd', 'Arm', 'Power',
           'Device', 'NvidiaDevice', 'AmdDevice', 'IntelDevice',
+           # Generic
+           'CPU64',


SO both CPU and Cpu ?

CPU64 = Cpu64('cpu64')

in archinfo

Agrred, CPU64 + Cpu64 seems like there shound't need both

one is the class the other an instance I agree it's pretty horrible will check

ok, done CPU64 -> ANYCPU

georgebisbas · 2024-06-19T14:46:53Z

devito/operator/operator.py

@@ -939,21 +955,23 @@ def _emit_apply_profiling(self, args):
            if metrics:
                perf("Global performance: [%s]" % ', '.join(metrics))

+            # Same as above, but excluding the setup phase, e.g. the CPU-GPU


So the data transfer time from device to host is not included ?
is it reported somewhere else?

wdym ? this is just an additional line -- a line that does not include the setup phase

I will cross-check the logs again

devito/operator/operator.py

devito/tools/abc.py

georgebisbas · 2024-06-19T14:56:18Z

devito/operator/operator.py

+                metrics.append("%.2f s" % fround(v.time))
+                metrics.append("%.2f GPts/s" % fround(v.gpointss))
+
+                perf("Global performance <w/o setup>: [%s]" % ', '.join(metrics))


I tried to run locally the PR and I do not see any diff in the log, should I ?

it should yes according to these changes

mloubout · 2024-06-21T14:49:37Z

devito/tools/abc.py

@@ -151,6 +151,22 @@ def __init__(self, a, b, c=4):

        return cls(*args, **kwargs)

+    def _rreplace(self, subs):


Why? this is just self._rebuild(**subs)

not only was it pretty similar if not identical to _rebuild, but also -- I just noted, sorry -- it appears to be completely unused at this stage. Probably something I coded during development which I eventually managed to avoid using, hence I'm dropping it for good!

EdCaunt · 2024-06-25T14:11:04Z

devito/ir/clusters/algorithms.py

@@ -433,7 +433,7 @@ def callback(self, clusters, prefix, seen=None):

            key = lambda i: i in prefix[:-1] or i in hs.loc_indices
            ispace = c.ispace.project(key)
-            # HaloTOuch are not parallel
+            # HaloTouch's are not parallel


"HaloTouches"

EdCaunt · 2024-06-25T14:46:14Z

devito/operator/operator.py

@@ -975,21 +993,12 @@ def lower_perfentry(v):

            if v.time <= 0.01:
                # Trim down the output for very fast sections
-                name = "%s%s<>" % (k.name, rank)
+                name = "%s%s" % (k.name, rank)


Could you define this first thing in the function? Would save the repeated line

good catch, OK

EdCaunt · 2024-06-25T15:42:25Z

devito/types/basic.py

+    def _signature(self):
+        """
+        The signature of an AbstractFunction is the set of fields that
+        makes it "compatible" with another AbstractFunction. The fact that


Might be worth defining "compatible" here

that's encoded in the first part of the sentence

In what sense? The current docstring doesn't help me understand what it means for one AbstractFunction to be compatible with another in this context

I will try to improve it in my next PR

FabioLuporini added the compiler label Jun 19, 2024

FabioLuporini requested review from mloubout and georgebisbas June 19, 2024 12:46

FabioLuporini commented Jun 19, 2024

View reviewed changes

devito/ir/stree/algorithms.py Show resolved Hide resolved

georgebisbas reviewed Jun 19, 2024

View reviewed changes

mloubout approved these changes Jun 21, 2024

View reviewed changes

FabioLuporini force-pushed the data-stream-improvs branch from b79cf67 to 5d9dfc6 Compare June 24, 2024 08:20

EdCaunt reviewed Jun 25, 2024

View reviewed changes

EdCaunt approved these changes Jun 26, 2024

View reviewed changes

FabioLuporini added 17 commits June 26, 2024 15:02

compiler: Fix GPU buffering

2dd165c

compiler: Optimize normalize_reductions_dense

3877506

api: Emit no-setup performance metrics

44f2c0b

compiler: Fixup profiling in presence of asynchronous streams

36941b9

api: Get rid of itershapes in verbose output log

5445974

compiler: Add IrregularFunctionInterface

dc814b2

compiler: Add ArgumentsMap.nbytes_avail_mapper

08d04f5

compiler: Add AbstractFunction._signature

e47ff07

compiler: Add rudimentary Reconstructable._replace

f66713a

compiler: Deduct pthreads from nthreads

bc3cb1d

compiler: Fixup buffering

5af85a6

compiler: Tweak generation of Sections

d9c1dce

compiler: Patch uxreplace's HaloScheme handler

43e9e50

arch: CPU64 -> ANYCPU

dd803d7

compiler: Drop unused Ops.replace

a8c9d86

compiler: Patch NThreads._arg_defaults

1c0aa52

compiler: Drop redundant code

4d08839

FabioLuporini force-pushed the data-stream-improvs branch from 7ab94e2 to 4d08839 Compare June 26, 2024 15:03

FabioLuporini merged commit 4e86b12 into master Jun 28, 2024
31 checks passed

FabioLuporini deleted the data-stream-improvs branch June 28, 2024 07:13

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

compiler: Improve data streaming backend #2390

compiler: Improve data streaming backend #2390

FabioLuporini commented Jun 19, 2024

codecov bot commented Jun 19, 2024 •

edited

Loading

georgebisbas Jun 19, 2024

georgebisbas Jun 19, 2024

mloubout Jun 21, 2024

FabioLuporini Jun 21, 2024

FabioLuporini Jun 24, 2024

georgebisbas Jun 19, 2024

FabioLuporini Jun 24, 2024

georgebisbas Jun 24, 2024

georgebisbas Jun 19, 2024

FabioLuporini Jun 21, 2024

mloubout Jun 21, 2024

FabioLuporini Jun 24, 2024

EdCaunt Jun 25, 2024

FabioLuporini Jun 26, 2024

EdCaunt Jun 25, 2024

FabioLuporini Jun 26, 2024

EdCaunt Jun 25, 2024

FabioLuporini Jun 26, 2024

EdCaunt Jun 26, 2024

FabioLuporini Jun 28, 2024

		@@ -151,6 +151,22 @@ def __init__(self, a, b, c=4):

		return cls(args, *kwargs)

		def _rreplace(self, subs):

compiler: Improve data streaming backend #2390

compiler: Improve data streaming backend #2390

Conversation

FabioLuporini commented Jun 19, 2024

codecov bot commented Jun 19, 2024 • edited Loading

Codecov Report

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

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

codecov bot commented Jun 19, 2024 •

edited

Loading