Skip to content

hexagon: add support for basic and extended Op profiling #22269

Merged
max-krasnyansky merged 19 commits into
ggml-org:masterfrom
qualcomm:hexagon-profiler
Apr 23, 2026
Merged

hexagon: add support for basic and extended Op profiling #22269
max-krasnyansky merged 19 commits into
ggml-org:masterfrom
qualcomm:hexagon-profiler

Conversation

@max-krasnyansky

Copy link
Copy Markdown
Member

Overview

The op-batching rewrite as part of #21705 disabled our original sync-profiler support.
This PR adds fully async profiler properly integrated with op-batching. The overall op queing logic has been updated to allocate additional shared buffers that (if enabled) the NPU now populates with the profiling data for each op.

The overall "user-interface" is the same as before where GGML_HEXAGON_PROFILE=1 enables Hexagon profiling output into the main GGML log. We now support additional modes: 1-basic (usec and cycles), 2-extended (usec, cycles + PMU). See docs/backends/snapdragon/README.md for details.

Additional information

This PR also includes a new post-processing tool that generates nice reports from the logs.
Here is an example:

M=../gguf/LFM2.5-1.2B-Instruct-Q4_0.gguf D=HTP0 PROF=2 ./scripts/snapdragon/adb/run-completion.sh -f ../sample_prompt_1024.txt -st -n 64 2>&1 | tee hex-log

./scripts/snapdragon/ggml-hexagon-profile.py hex-log

# Profile Report (Top 100 Ops sorted by max-usec)

| Op             | Dims                                     | DTypes                       | Max usec | Avg usec | Max Cycles | Avg Cycles  | Count |
| -------------- | ---------------------------------------- | ---------------------------- | -------- | -------- | ---------- | ----------- | ----- |
| MUL_MAT        | 8192:2048 x 8192:255 -> 2048:255         | q4_0 x f32 -> f32            | 6153     | 6126.27  | 12998214   | 12940843.53 | 15    |
| MUL_MAT        | 2048:8192 x 2048:255 -> 8192:255         | q4_0 x f32 -> f32            | 5493     | 5450.40  | 11602839   | 11513942.20 | 30    |
| MUL_MAT        | 8192:2048 x 8192:256 -> 2048:256         | q4_0 x f32 -> f32            | 5461     | 5052.17  | 11536246   | 10672342.00 | 30    |

Requirements

  • I have read and agree with the contributing guidelines
  • AI usage disclosure: Not for the core C/C++ code; Used Gemini to generate initial version of the post-proc tool

@max-krasnyansky
max-krasnyansky requested a review from a team as a code owner April 23, 2026 01:28
@github-actions github-actions Bot added documentation Improvements or additions to documentation script Script related python python script changes ggml changes relating to the ggml tensor library for machine learning Hexagon labels Apr 23, 2026
@max-krasnyansky

Copy link
Copy Markdown
Member Author

@ggml-org/maintainers need the second approval here (any input is welcome too of course).

Comment thread ggml/src/ggml-hexagon/ggml-hexagon.cpp
Comment thread ggml/src/ggml-hexagon/htp/htp-ops.h Outdated
Comment thread scripts/snapdragon/ggml-hexagon-profile.py Outdated
max-krasnyansky and others added 2 commits April 23, 2026 14:09
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
@max-krasnyansky

max-krasnyansky commented Apr 23, 2026

Copy link
Copy Markdown
Member Author

@lhez @CISC sorry, need re-approvals after applying suggestions.
That new feature (ie making approvals stale) is a bit of a time waster, makes sense for general contribs but @ggerganov maybe it can be disabled if the PRs authors are maintainers?

@max-krasnyansky
max-krasnyansky merged commit 5d2b52d into ggml-org:master Apr 23, 2026
5 of 50 checks passed
IntelNav pushed a commit to IntelNav/llama.cpp that referenced this pull request Apr 29, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
IntelNav pushed a commit to IntelNav/llama.cpp that referenced this pull request Apr 29, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
samuraieng pushed a commit to samuraieng/llama.cpp that referenced this pull request May 6, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
ljubomirj pushed a commit to ljubomirj/llama.cpp that referenced this pull request May 6, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
meh pushed a commit to meh/llama.cpp that referenced this pull request May 10, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
my-other-github-account pushed a commit to my-other-github-account/llama.cpp that referenced this pull request May 15, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
my-other-github-account pushed a commit to my-other-github-account/llama.cpp that referenced this pull request May 15, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
baramofme pushed a commit to baramofme/llama-cpp-turboquant that referenced this pull request May 23, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
fewtarius pushed a commit to fewtarius/CachyLLama that referenced this pull request May 30, 2026
)

* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
MrLordCat referenced this pull request in MrLordCat/llama.cpp-with-GUI Jul 16, 2026
* hexagon: restore HTP_OPMASK_QUEUE

* hexagon: honor OPMASK_SKIP_COMPUTE in hmx-matmul

* hex-prof: restore op profiling

* hex-prof: enable PMU

* hexagon: simplify and improve op-queuing with full profiling support

Add separate profile descriptors.

* hexagon: remove opsync and rename opmask into opstage

opsync is no longer needed since the profiler is fully async now.
opmask name was confusing and opstage is more accurate.

* hexagon: refactor opbatch queue handling

* hexagon: add iface hooks for enabling profiler from the host

Also move all the PMU setup stuff out of the hex-utils since it's not inteded for normal use.

* hexagon: make profiler mode configurable

On older devices getting PMU counters is expensive so it's now optional.

* hexagon: add support for setting profiler pmu events from env

* hexagon: simplify profiler output (no need to print buffs, etc)

* hexagon: simplify pmu counter formating

* hexagon: add a simple profile post-proc tool

* hex-prof: add support for reading logs from stdin

* hexagon: document GGML_HEXAGON_PROFILE

* hex-prof: update default width for dims field

* hex-prof: fix linter warnings and errors

* Update ggml/src/ggml-hexagon/htp/htp-ops.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update scripts/snapdragon/ggml-hexagon-profile.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation ggml changes relating to the ggml tensor library for machine learning Hexagon python python script changes script Script related

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants