Feature: mark a Func as no_profiling, to prevent injection of profiling. (2nd implementation) #8143

mcourteaux · 2024-03-07T13:02:51Z

Replaces #8136 with a cleaner implementation.

@abadams I took your suggestion, and this implementation is much better. Very small changes, and the issue I reported in the previous one is gone. The trick with the "environment" was a good lesson!

Additionally, I managed to also clean up the report by having the stack-allocated Allocate/Free nodes also have their allocation contribute to the enclosing Func. Doing that changed my profiling report from (w_0, w_1, w_2, max_channel, lum, lum_pow3, r, g, b all marked with no_profiling(), but still producing lines in the report):

neonraw_highlight_ratio_extractor
 total time: 1338.695068 ms  samples: 1262  runs: 48  time/run: 27.889481 ms
 average threads used: 2.657686
 heap allocations: 0  peak heap usage: 0 bytes
    halide_malloc:                0.000ms ( 0.0%)   threads: 0.000 
    halide_free:                  0.000ms ( 0.0%)   threads: 0.000 
    ratio_map:                   21.282ms (76.3%)   threads: 2.663 
    repeat_edge:                  2.764ms ( 9.9%)   threads: 3.248  stack: 3960
    max_field_hpass:              0.683ms ( 2.4%)   threads: 2.967  stack: 3840
    max_field:                    0.352ms ( 1.2%)   threads: 2.125  stack: 3072
    min_field_hpass:              0.507ms ( 1.8%)   threads: 2.826  stack: 1280
    max_channel:                  0.000ms ( 0.0%)   threads: 0.000  stack: 132
    min_field:                    0.088ms ( 0.3%)   threads: 1.250  stack: 1024
    b:                            0.000ms ( 0.0%)   threads: 0.000  stack: 64
    g:                            0.000ms ( 0.0%)   threads: 0.000  stack: 64
    r:                            0.000ms ( 0.0%)   threads: 0.000  stack: 64
    lum:                          0.000ms ( 0.0%)   threads: 0.000  stack: 64
    lum_pow3:                     0.000ms ( 0.0%)   threads: 0.000  stack: 64
    lum4:                         2.210ms ( 7.9%)   threads: 1.870  stack: 64
    w_0:                          0.000ms ( 0.0%)   threads: 0.000  stack: 64
    w_1:                          0.000ms ( 0.0%)   threads: 0.000  stack: 64
    w_2:                          0.000ms ( 0.0%)   threads: 0.000  stack: 64

to this (note how the stack size contributions (132 + 8 * 64 = 644) moved up into the enclosing Funcs:

min_field_hpass: 1412 - 1280 = 132 bytes extra,
ratio_map 384 - 0 = 384 bytes extra

The total difference is less than 644 bytes, because the lifetimes of some of the variables were actually fully-disjunct, causing them to not actually have a peak stack usage of that much.):

neonraw_highlight_ratio_extractor
 total time: 1524.809326 ms  samples: 1416  runs: 48  time/run: 31.766861 ms
 average threads used: 4.375706
 heap allocations: 0  peak heap usage: 0 bytes
    halide_malloc:                0.000ms ( 0.0%)   threads: 0.000 
    halide_free:                  0.000ms ( 0.0%)   threads: 0.000 
    ratio_map:                   24.556ms (77.3%)   threads: 4.067  stack: 384
    repeat_edge:                  3.719ms (11.7%)   threads: 5.716  stack: 3960
    max_field_hpass:              0.420ms ( 1.3%)   threads: 3.000  stack: 3840
    max_field:                    0.336ms ( 1.0%)   threads: 5.666  stack: 3072
    min_field_hpass:              0.575ms ( 1.8%)   threads: 3.846  stack: 1412
    min_field:                    0.088ms ( 0.2%)   threads: 8.500  stack: 1024
    lum4:                         2.069ms ( 6.5%)   threads: 5.813  stack: 64

…oop) Func should not be profiled.

abadams · 2024-03-07T19:01:23Z

src/Profiling.cpp

+ }
+ }
+
+ const Function *lookup_function(const string &name) const {


Function is already a pointer type, so this should probably just return a Function by value. Alternatively you could return a const Function &, but I'm not sure how to satisfy the compiler in the case where you need to return something valid after the internal_error. Maybe just return env.begin()->second?

I first had a reference, but I in case the lookup fails, I need to satisfy the compiler and give it a nullptr. Idk if we can easily create a nullptr-Function& instead of a Function* being nullptr?

I suggest satisfying the compiler by returning the first Function in the environment: env.begin()->second.

There is a default constructor I see. So for this unreachable code, a simple return {}; looks a bit nicer I think?

Isn't that returning a reference to a local? Given that it's unreachable, whatever makes the compiler shut up is fine.

Isn't that returning a reference to a local?

AFAICT, no. It's invoking the default constructor (the one literally defined with = default;), which just initialized the FunctionPtr inside to be default-constructed, meaning that there are just some pointers put to nullptr I assume.

abadams · 2024-03-07T19:01:59Z

src/halide_ir.fbs

@@ -713,6 +713,7 @@ table Func {
 trace_stores: bool = false;
 trace_realizations: bool = false;
 trace_tags: [string];
+ no_profiling: bool = false;


Probably need to bump the serialization minor version number

I'm actually clueless where to find that number.

I'm a bit confused too. @TH3CHARLie ?

it gets refactored earlier to here:

Halide/src/halide_ir.fbs

Line 9 in 8cc4f02

enum SerializationVersionMajor: int {

@TH3CHARLie looking at those inline comments, suggest that I shouldn't touch anything, as we are still at "unstable"? Do I bump the patch version instead?

yeah I believe so

@abadams @TH3CHARLie It all looks suspicious, as the Patch version was at 0 (zero). Serialization version 18.0.0 seems suspicious, but maybe is what we are after? I just bumped it to 18.0.1.

It was at zero since we recently had a release and introduced this version numbering system and hasn't added anything new to the format, I don't understand what's suspicious here.

Makes sense! Thanks!

…of the serialization.

mcourteaux · 2024-03-12T21:24:43Z

AFAICT the failing build is unrelated, right? This might conflict with abadams/per_instance_profiling, in which case I'm happy to adapt this PR after we merge in Andrew's work.

abadams · 2024-03-12T23:56:34Z

Yes, the failure is unrelated (fixed by #8148)

mcourteaux added 3 commits March 7, 2024 13:27

Small feature to allow you to specify that a (typically small inner l…

d1ca091

…oop) Func should not be profiled.

Simplified the tuple name handling.

e3643c0

Optimize tuple name normalization in Profiling.cpp

d011e43

mcourteaux changed the title ~~Small feature to allow you to specify that a (typically small inner loop) Func should not be profiled.~~ Feature: mark a Func as no_profiling, to prevent injection of profiling. Mar 7, 2024

mcourteaux changed the title ~~Feature: mark a Func as no_profiling, to prevent injection of profiling.~~ Feature: mark a Func as no_profiling, to prevent injection of profiling. (2nd implementation) Mar 7, 2024

mcourteaux mentioned this pull request Mar 7, 2024

Feature: mark a Func as no_profiling, to prevent injection of profiling. #8136

Closed

Clang-format

25f9990

mcourteaux force-pushed the func-no-profile-2 branch from c030bb1 to 25f9990 Compare March 7, 2024 13:23

steven-johnson requested a review from abadams March 7, 2024 17:23

abadams reviewed Mar 7, 2024

View reviewed changes

Feedback on Function already being a pointer. Bump the Patch version …

b8bda38

…of the serialization.

abadams requested a review from halidebuildbots March 9, 2024 22:33

abadams approved these changes Mar 12, 2024

View reviewed changes

abadams merged commit 4988ab5 into halide:main Mar 12, 2024
18 of 19 checks passed

abadams added the release_notes For changes that may warrant a note in README for official releases. label Mar 13, 2024

BrewTestBot mentioned this pull request Jul 17, 2024

halide 18.0.0 Homebrew/homebrew-core#177657

Closed

1 task

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Feature: mark a Func as no_profiling, to prevent injection of profiling. (2nd implementation) #8143

Feature: mark a Func as no_profiling, to prevent injection of profiling. (2nd implementation) #8143

mcourteaux commented Mar 7, 2024 •

edited

Loading

abadams Mar 7, 2024

mcourteaux Mar 7, 2024

abadams Mar 7, 2024

mcourteaux Mar 7, 2024

abadams Mar 8, 2024

mcourteaux Mar 8, 2024 •

edited

Loading

abadams Mar 7, 2024

mcourteaux Mar 7, 2024

abadams Mar 8, 2024

TH3CHARLie Mar 8, 2024

mcourteaux Mar 8, 2024

TH3CHARLie Mar 8, 2024

mcourteaux Mar 8, 2024

TH3CHARLie Mar 8, 2024

mcourteaux Mar 9, 2024

mcourteaux commented Mar 12, 2024

abadams commented Mar 12, 2024

Feature: mark a Func as no_profiling, to prevent injection of profiling. (2nd implementation) #8143

Feature: mark a Func as no_profiling, to prevent injection of profiling. (2nd implementation) #8143

Conversation

mcourteaux commented Mar 7, 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

mcourteaux Mar 8, 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

mcourteaux commented Mar 12, 2024

abadams commented Mar 12, 2024

mcourteaux commented Mar 7, 2024 •

edited

Loading

mcourteaux Mar 8, 2024 •

edited

Loading