Skip to content

Commit 398ae1e

Browse files
trivedivivektrivedivivek
committed
Update on "[ET-VK] Removing unnecessary and redundant members from VulkanBuffer and ParamsBuffer."
This diff removes unnecessary and redundant members from VulkanBuffer and ParamsBuffer. Differential Revision: [D66557456](https://our.internmc.facebook.com/intern/diff/D66557456/) [ghstack-poisoned]
2 parents 278bd83 + 0db32f6 commit 398ae1e

File tree

40 files changed

+1727
-264
lines changed

40 files changed

+1727
-264
lines changed

‎.github/scripts/extract_benchmark_results.py

Lines changed: 104 additions & 49 deletions
Original file line numberDiff line numberDiff line change
@@ -310,6 +310,7 @@ def transform(
310310
workflow_run_attempt: int,
311311
job_name: str,
312312
job_id: int,
313+
schema_version: str,
313314
) -> List:
314315
"""
315316
Transform the benchmark results into the format writable into the benchmark database
@@ -319,45 +320,91 @@ def transform(
319320
for r in benchmark_results:
320321
r["deviceInfo"]["device"] = job_name
321322

322-
# TODO (huydhn): This is the current schema of the database oss_ci_benchmark_v2,
323-
# and I'm trying to fit ET benchmark results into it, which is kind of awkward.
324-
# However, the schema is going to be updated soon
325-
return [
326-
{
327-
# GH-info to identify where the benchmark is run
328-
"repo": repo,
329-
"head_branch": head_branch,
330-
"workflow_id": workflow_run_id,
331-
"run_attempt": workflow_run_attempt,
332-
"job_id": job_id,
333-
# The model
334-
"name": f"{r['benchmarkModel']['name']} {r['benchmarkModel'].get('backend', '')}".strip(),
335-
"dtype": (
336-
r["benchmarkModel"]["quantization"]
337-
if r["benchmarkModel"]["quantization"]
338-
else "unknown"
339-
),
340-
# The metric value
341-
"metric": r["metric"],
342-
"actual": r["actualValue"],
343-
"target": r["targetValue"],
344-
# The device
345-
"device": r["deviceInfo"]["device"],
346-
"arch": r["deviceInfo"].get("os", ""),
347-
# Not used here, just set it to something unique here
348-
"filename": workflow_name,
349-
"test_name": app_type,
350-
"runner": job_name,
351-
}
352-
for r in benchmark_results
353-
]
323+
if schema_version == "v2":
324+
# TODO (huydhn): Clean up this branch after ExecuTorch dashboard migrates to v3
325+
return [
326+
{
327+
# GH-info to identify where the benchmark is run
328+
"repo": repo,
329+
"head_branch": head_branch,
330+
"workflow_id": workflow_run_id,
331+
"run_attempt": workflow_run_attempt,
332+
"job_id": job_id,
333+
# The model
334+
"name": f"{r['benchmarkModel']['name']} {r['benchmarkModel'].get('backend', '')}".strip(),
335+
"dtype": (
336+
r["benchmarkModel"]["quantization"]
337+
if r["benchmarkModel"]["quantization"]
338+
else "unknown"
339+
),
340+
# The metric value
341+
"metric": r["metric"],
342+
"actual": r["actualValue"],
343+
"target": r["targetValue"],
344+
# The device
345+
"device": r["deviceInfo"]["device"],
346+
"arch": r["deviceInfo"].get("os", ""),
347+
# Not used here, just set it to something unique here
348+
"filename": workflow_name,
349+
"test_name": app_type,
350+
"runner": job_name,
351+
}
352+
for r in benchmark_results
353+
]
354+
elif schema_version == "v3":
355+
quantization = (
356+
r["benchmarkModel"]["quantization"]
357+
if r["benchmarkModel"]["quantization"]
358+
else "unknown"
359+
)
360+
# From https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database
361+
return [
362+
{
363+
"benchmark": {
364+
"name": "ExecuTorch",
365+
"mode": "inference",
366+
"dtype": quantization,
367+
"extra_info": {
368+
"app_type": app_type,
369+
},
370+
},
371+
"model": {
372+
"name": r["benchmarkModel"]["name"],
373+
"type": "OSS model",
374+
"backend": r["benchmarkModel"].get("backend", ""),
375+
"extra_info": {
376+
"quantization": quantization,
377+
},
378+
},
379+
"metric": {
380+
"name": r["metric"],
381+
"benchmark_values": [r["actualValue"]],
382+
"target_value": r["targetValue"],
383+
"extra_info": {
384+
"method": r.get("method", ""),
385+
},
386+
},
387+
"runners": [
388+
{
389+
"name": r["deviceInfo"]["device"],
390+
"type": r["deviceInfo"]["os"],
391+
"avail_mem_in_gb": r["deviceInfo"].get("availMem", ""),
392+
"total_mem_in_gb": r["deviceInfo"].get("totalMem", ""),
393+
}
394+
],
395+
}
396+
for r in benchmark_results
397+
]
354398

355399

356400
def main() -> None:
357401
args = parse_args()
358402

359-
# Across all devices
360-
all_benchmark_results = []
403+
# Across all devices, keeping both schemas for now until ExecuTorch dashboard migrates to v3
404+
all_benchmark_results = {
405+
"v2": [],
406+
"v3": [],
407+
}
361408

362409
with open(args.artifacts) as f:
363410
for artifact in json.load(f):
@@ -384,23 +431,31 @@ def main() -> None:
384431
)
385432

386433
if benchmark_results:
387-
benchmark_results = transform(
388-
app_type,
389-
benchmark_results,
390-
args.repo,
391-
args.head_branch,
392-
args.workflow_name,
393-
args.workflow_run_id,
394-
args.workflow_run_attempt,
395-
job_name,
396-
extract_job_id(args.artifacts),
397-
)
398-
all_benchmark_results.extend(benchmark_results)
434+
for schema in all_benchmark_results.keys():
435+
results = transform(
436+
app_type,
437+
benchmark_results,
438+
args.repo,
439+
args.head_branch,
440+
args.workflow_name,
441+
args.workflow_run_id,
442+
args.workflow_run_attempt,
443+
job_name,
444+
extract_job_id(args.artifacts),
445+
schema,
446+
)
447+
all_benchmark_results[schema].extend(results)
448+
449+
for schema in all_benchmark_results.keys():
450+
if not all_benchmark_results.get(schema):
451+
continue
452+
453+
output_dir = os.path.join(args.output_dir, schema)
454+
os.mkdir(output_dir)
399455

400-
if all_benchmark_results:
401456
output_file = os.path.basename(args.artifacts)
402-
with open(f"{args.output_dir}/{output_file}", "w") as f:
403-
json.dump(all_benchmark_results, f)
457+
with open(f"{output_dir}/{output_file}", "w") as f:
458+
json.dump(all_benchmark_results[schema], f)
404459

405460

406461
if __name__ == "__main__":

‎.github/workflows/android-perf.yml

Lines changed: 17 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -298,15 +298,25 @@ jobs:
298298
--workflow-run-attempt ${{ github.run_attempt }}
299299
done
300300
301-
ls -lah benchmark-results
302-
303-
for BENCHMARK_RESULTS in benchmark-results/*.json; do
304-
cat "${BENCHMARK_RESULTS}"
305-
echo
301+
for SCHEMA in v2 v3; do
302+
for BENCHMARK_RESULTS in benchmark-results/"${SCHEMA}"/*.json; do
303+
cat "${BENCHMARK_RESULTS}"
304+
echo
305+
done
306306
done
307307
308-
- name: Upload the benchmark results
308+
# TODO (huydhn): Remove v2 schema once the benchmark dashboard finishes the migration
309+
- name: Upload the benchmark results (v2)
310+
uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
311+
with:
312+
benchmark-results-dir: benchmark-results/v2
313+
dry-run: false
314+
schema-version: v2
315+
316+
- name: Upload the benchmark results (v3)
309317
uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
310318
with:
311-
benchmark-results-dir: 'benchmark-results'
319+
benchmark-results-dir: benchmark-results/v3
312320
dry-run: false
321+
schema-version: v3
322+
github-token: ${{ secrets.GITHUB_TOKEN }}

‎.github/workflows/apple-perf.yml

Lines changed: 17 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -372,15 +372,25 @@ jobs:
372372
--workflow-run-attempt ${{ github.run_attempt }}
373373
done
374374
375-
ls -lah benchmark-results
376-
377-
for BENCHMARK_RESULTS in benchmark-results/*.json; do
378-
cat "${BENCHMARK_RESULTS}"
379-
echo
375+
for SCHEMA in v2 v3; do
376+
for BENCHMARK_RESULTS in benchmark-results/"${SCHEMA}"/*.json; do
377+
cat "${BENCHMARK_RESULTS}"
378+
echo
379+
done
380380
done
381381
382-
- name: Upload the benchmark results
382+
# TODO (huydhn): Remove v2 schema once the benchmark dashboard finishes the migration
383+
- name: Upload the benchmark results (v2)
384+
uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
385+
with:
386+
benchmark-results-dir: benchmark-results/v2
387+
dry-run: false
388+
schema-version: v2
389+
390+
- name: Upload the benchmark results (v3)
383391
uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
384392
with:
385-
benchmark-results-dir: 'benchmark-results'
393+
benchmark-results-dir: benchmark-results/v3
386394
dry-run: false
395+
schema-version: v3
396+
github-token: ${{ secrets.GITHUB_TOKEN }}

‎CMakeLists.txt

Lines changed: 15 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -56,6 +56,21 @@ if(NOT CMAKE_BUILD_TYPE)
5656
set(CMAKE_BUILD_TYPE Debug)
5757
endif()
5858

59+
# Setup RPATH.
60+
# See https://gitlab.kitware.com/cmake/community/-/wikis/doc/cmake/RPATH-handling
61+
# Use separate rpaths during build and install phases
62+
set(CMAKE_SKIP_BUILD_RPATH OFF)
63+
# Don't use the install-rpath during the build phase
64+
set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
65+
# Automatically add all linked folders that are NOT in the build directory to
66+
# the rpath (per library?)
67+
# TODO: Doesn't work for us right now because we are not installing .so's into the
68+
# correct locations. For example we have libcustom_ops_aot_lib.so depending on
69+
# _portable_lib.so, which was eventually put under <site-packages>/executorch/extension/pybindings/
70+
# but this rpath is not automatically added because at build time it seems `portable_lib`
71+
# is being built under the same directory, so no extra rpath is being added. To
72+
# properly fix this we need to install `portable_lib` into the correct path.
73+
set(CMAKE_INSTALL_RPATH_USE_LINK_PATH ON)
5974
# ------------------------------ OPTIONS -------------------------------------
6075
# WARNING: Please don't add example specific options in this CMakeLists.txt.
6176
# Instead please use `find_package(executorch REQUIRED)` in the example
@@ -682,22 +697,6 @@ if(EXECUTORCH_BUILD_PTHREADPOOL
682697
endif()
683698

684699
if(EXECUTORCH_BUILD_PYBIND)
685-
# Setup RPATH.
686-
# See https://gitlab.kitware.com/cmake/community/-/wikis/doc/cmake/RPATH-handling
687-
if(APPLE)
688-
set(CMAKE_MACOSX_RPATH ON)
689-
set(_rpath_portable_origin "@loader_path")
690-
else()
691-
set(_rpath_portable_origin $ORIGIN)
692-
endif(APPLE)
693-
# Use separate rpaths during build and install phases
694-
set(CMAKE_SKIP_BUILD_RPATH FALSE)
695-
# Don't use the install-rpath during the build phase
696-
set(CMAKE_BUILD_WITH_INSTALL_RPATH FALSE)
697-
set(CMAKE_INSTALL_RPATH "${_rpath_portable_origin}")
698-
# Automatically add all linked folders that are NOT in the build directory to
699-
# the rpath (per library?)
700-
set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
701700
add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/third-party/pybind11)
702701

703702
if(NOT EXECUTORCH_BUILD_EXTENSION_DATA_LOADER)

‎backends/cadence/aot/compiler.py

Lines changed: 22 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -28,6 +28,7 @@
2828
to_edge,
2929
)
3030
from executorch.exir.pass_base import PassResult
31+
from torch._inductor.decomposition import remove_decompositions
3132
from torch.ao.quantization.pt2e.export_utils import model_is_exported
3233
from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
3334

@@ -58,16 +59,33 @@ def convert_pt2(
5859
Returns a GraphModule with the converted model.
5960
"""
6061

62+
# Get default decompositions
63+
decomp_table = torch.export.default_decompositions()
64+
# Select ops to keep
65+
ops_to_keep = [
66+
torch.ops.aten.conv1d.default,
67+
torch.ops.aten.conv2d.default,
68+
torch.ops.aten.layer_norm.default,
69+
torch.ops.aten.linear.default,
70+
torch.ops.aten.matmul.default,
71+
]
72+
# Remove decompositions for the ops we want to keep
73+
# pyre-fixme[6]: For 1st argument expected `Dict[typing.Callable[..., typing.Any
74+
remove_decompositions(decomp_table, ops_to_keep)
6175
# Export with dynamo
62-
model_gm = torch.export.export_for_training(model, inputs).module()
76+
model_gm = (
77+
torch.export.export_for_training(model, inputs)
78+
.run_decompositions(decomp_table)
79+
.module()
80+
)
6381

64-
if model_gm_has_SDPA(model_gm): # pyre-fixme[6]
82+
if model_gm_has_SDPA(model_gm):
6583
# Decompose SDPA
66-
DecomposeScaledDotProductAttention(False)(model_gm) # pyre-fixme[6]
84+
DecomposeScaledDotProductAttention(False)(model_gm)
6785

6886
# Swap _safe_softmax with _softmax (see https://github.com/pytorch/pytorch/pull/133882
6987
# for details).
70-
result = ReplaceSafeSoftmaxWithSoftmax()(model_gm) # pyre-fixme[6]
88+
result = ReplaceSafeSoftmaxWithSoftmax()(model_gm)
7189
assert result is not None
7290
model_gm = result.graph_module
7391

‎backends/cadence/hifi/kernels/kernels.h

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -92,9 +92,9 @@ extern "C" WORD32 xa_nn_elm_mul_broadcast_4D_f32xf32_f32(
9292
const WORD32* const p_inp2_shape);
9393

9494
extern "C" void xa_nn_elm_pow_f32(
95-
FLOAT32* restrict z,
96-
const FLOAT32* restrict x,
97-
const FLOAT32* restrict y,
95+
FLOAT32* __restrict__ z,
96+
const FLOAT32* __restrict__ x,
97+
const FLOAT32* __restrict__ y,
9898
WORD32 N);
9999

100100
extern "C" WORD32 xa_nn_elm_where_f32xf32_f32(

‎backends/cadence/hifi/third-party/nnlib/xa_nn_elm_minimum_maximum_f32.c

Lines changed: 7 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -19,12 +19,12 @@
1919
* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
2020
2121
******************************************************************************/
22-
#include "nnlib-hifi4/xa_nnlib/include/xa_type_def.h"
23-
#include "nnlib-hifi4/xa_nnlib/algo/common/include/xa_nnlib_common_fpu.h"
24-
#include "nnlib-hifi4/xa_nnlib/algo/common/include/xa_nn_common.h"
25-
#include "nnlib-hifi4/xa_nnlib/algo/common/include/xa_nnlib_err_chk.h"
26-
#include "nnlib-hifi4/xa_nnlib/algo/kernels/basic/hifi4/xa_nn_basic_state.h"
27-
#include "nnlib-hifi4/xa_nnlib/include/nnlib/xa_nnlib_kernels_api.h"
22+
#include "xa_type_def.h"
23+
#include "xa_nnlib_common_fpu.h"
24+
#include "xa_nn_common.h"
25+
#include "xa_nnlib_err_chk.h"
26+
// #include "xa_nn_basic_state.h"
27+
#include "xa_nnlib_kernels_api.h"
2828

2929
#if !HAVE_VFPU
3030
DISCARD_FUN_FOR_NONVOID_RETURN(
@@ -844,4 +844,4 @@ WORD32 xa_nn_elm_minimum_broadcast_4D_f32xf32_f32(FLOAT32 * __restrict__ p_out,
844844
}
845845
return 0;
846846
}
847-
#endif
847+
#endif

‎backends/test/README.md

Whitespace-only changes.

‎backends/test/TARGETS

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,8 @@
1+
# Any targets that should be shared between fbcode and xplat must be defined in
2+
# targets.bzl. This file can contain fbcode-only targets.
3+
4+
load(":targets.bzl", "define_common_targets")
5+
6+
oncall("executorch")
7+
8+
define_common_targets(is_fbcode = True)

0 commit comments

Comments
 (0)