Skip to content

Commit bb130aa

Browse files
WangJialei-AWangJialei-A
committed
introduce benchgc for correctness check
1 parent 8500551 commit bb130aa

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

49 files changed

+5409
-0
lines changed

‎.github/workflows/build.yml

Lines changed: 13 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -45,3 +45,16 @@ jobs:
4545
- name: Test
4646
run: |
4747
cmake --build build --target gc-check
48+
49+
- name: Build and install benchgc
50+
working-directory: build
51+
run: |
52+
touch python_packages/gc_mlir_core/gc_mlir/extras/__init__.py
53+
ninja benchgc
54+
pip uninstall -y benchgc || true
55+
pip install test/benchgc/dist/benchgc-*.whl
56+
- name: Correctness Test
57+
env:
58+
LD_PRELOAD: /lib/x86_64-linux-gnu/libomp5.so
59+
run: |
60+
scripts/correctness.sh

‎.github/workflows/style.yml

Lines changed: 21 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -22,3 +22,24 @@ jobs:
2222
run: |
2323
clang-format --version
2424
find . -name *.cpp -or -name *.hpp | xargs clang-format --dry-run --Werror -style=file
25+
26+
python_format:
27+
runs-on: ubuntu-latest
28+
steps:
29+
- name: checkout base version
30+
uses: actions/checkout@v4
31+
with:
32+
fetch-depth: 100
33+
ref: ${{ github.event.pull_request.base.sha }}
34+
35+
- name: checkout head version
36+
uses: actions/checkout@v4
37+
with:
38+
fetch-depth: 100
39+
ref: ${{ github.event.pull_request.head.sha }}
40+
41+
- name: install darker
42+
run: "python3 -m pip install darker"
43+
44+
- name: check python format
45+
run: "python3 -m darker --check --diff -r `git merge-base ${{ github.event.pull_request.base.sha }} ${{ github.event.pull_request.head.sha }}`...HEAD ."

‎.gitignore

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -3,3 +3,4 @@
33
build/
44
externals/
55
compile_commands.json
6+
__pycache__

‎CMakeLists.txt

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -48,6 +48,7 @@ option(GC_ENABLE_OPT "Build gc-opt" ${GC_ENABLE_TOOLS})
4848
option(GC_ENABLE_IMEX "Enable Intel® Extension for MLIR" OFF)
4949
option(GC_ENABLE_BINDINGS_PYTHON "Enable Graph Complier Python Binding" ON)
5050
option(GC_DEV_LINK_LLVM_DYLIB "Link dynamic libraries of LLVM and MLIR. For developers only. Do not use it in packing the library." OFF)
51+
option(GC_BENCH_ENABLE "Build benchgc." ON)
5152

5253
if(GC_ENABLE_LEGACY)
5354
add_subdirectory(legacy/core)

‎scripts/correctness.sh

Lines changed: 97 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,97 @@
1+
#! /bin/bash
2+
3+
export CASE_DIR=$(pwd)/test/benchgc/cases
4+
5+
FAIL=0
6+
set -e
7+
8+
# misc
9+
python3 -m benchgc --verbose 0 --driver linalg --case fill --md 0:f32 --md 1:32x4096xf32 --cmp 1:P:0:0 || FAIL=1
10+
python3 -m benchgc --verbose 0 --driver linalg --case copy --md 0:1024x1024xf32 --md 1:1024x1024xbf16 || FAIL=1
11+
python3 -m benchgc --verbose 0 --driver linalg --case broadcast --md 0:1024xf32 --md 1:2x32x1024xf32 --dimensions=0 --dimensions=1 || FAIL=1
12+
13+
# matmul
14+
python3 -m benchgc --verbose 0 --driver linalg --case batch_matmul --md 0:16x512x64xf32 --md 1:16x64x32xf32 --md 2:16x512x32xf32 || FAIL=1
15+
python3 -m benchgc --verbose 0 --driver linalg --case batch_matmul_transpose_a --md 0:16x512x64xf32 --md 1:16x512x32xf32 --md 2:16x64x32xf32 || FAIL=1
16+
python3 -m benchgc --verbose 0 --driver linalg --case batch_matmul_transpose_b --md 0:16x512x64xf32 --md 1:16x128x64xf32 --md 2:16x512x128xf32 || FAIL=1
17+
python3 -m benchgc --verbose 0 --driver linalg --case batch_matvec --md 0:16x512x64xf32 --md 1:16x64xf32 --md 2:16x512xf32 || FAIL=1
18+
python3 -m benchgc --verbose 0 --driver linalg --case batch_mmt4d --md 0:4x4x8x4x2xf32 --md 1:4x8x8x4x2xf32 --md 2:4x4x8x4x4xf32 || FAIL=1
19+
python3 -m benchgc --verbose 0 --driver linalg --case batch_reduce_matmul --md 0:16x512x64xf32 --md 1:16x64x32xf32 --md 2:512x32xf32 || FAIL=1
20+
python3 -m benchgc --verbose 0 --driver linalg --case batch_vecmat --md 0:16x64xf32 --md 1:16x64x512xf32 --md 2:16x512xf32 || FAIL=1
21+
python3 -m benchgc --verbose 0 --driver linalg --case dot --md 0:4096xf32 --md 1:4096xf32 --md 2:0xf32 || FAIL=1
22+
python3 -m benchgc --verbose 0 --driver linalg --case matmul --md 0:1024x512xf32 --md 1:512x512xf32 --md 2:1024x512xf32 --cast cast_signed || FAIL=1
23+
python3 -m benchgc --verbose 0 --driver linalg --case matmul_transpose_a --md 0:1024x512xf32 --md 1:1024x512xf32 --md 2:512x512xf32 --cast cast_signed || FAIL=1
24+
python3 -m benchgc --verbose 0 --driver linalg --case matmul_transpose_b --md 0:1024x512xf32 --md 1:1024x512xf32 --md 2:1024x1024xf32 --cast cast_signed || FAIL=1
25+
python3 -m benchgc --verbose 0 --driver linalg --case matvec --md 0:512x64xf32 --md 1:64xf32 --md 2:512xf32 || FAIL=1
26+
python3 -m benchgc --verbose 0 --driver linalg --case mmt4d --md 0:4x8x4x2xf32 --md 1:8x8x4x2xf32 --md 2:4x8x4x4xf32 || FAIL=1
27+
python3 -m benchgc --verbose 0 --driver linalg --case vecmat --md 0:512xf32 --md 1:512x64xf32 --md 2:64xf32 || FAIL=1
28+
29+
# binary
30+
python3 -m benchgc --verbose 0 --driver linalg --case add --md 0:1x32x4096xf32 --md 1:1x32x4096xf32 --md 2:1x32x4096xf32 || FAIL=1
31+
python3 -m benchgc --verbose 0 --driver linalg --case sub --md 0:1x32x4096xf32 --md 1:1x32x4096xf32 --md 2:1x32x4096xf32 || FAIL=1
32+
python3 -m benchgc --verbose 0 --driver linalg --case mul --md 0:1x32x4096xf32 --md 1:1x32x4096xf32 --md 2:1x32x4096xf32 || FAIL=1
33+
python3 -m benchgc --verbose 0 --driver linalg --case div --md 0:1x32x4096xf32 --md 1:1x32x4096xf32 --md 2:1x32x4096xf32 || FAIL=1
34+
python3 -m benchgc --verbose 0 --driver linalg --case max --md 0:1024x1024xf32 --md 1:1024x1024xf32 --md 2:1024x1024xf32 || FAIL=1
35+
python3 -m benchgc --verbose 0 --driver linalg --case min --md 0:1024x1024xf32 --md 1:1024x1024xf32 --md 2:1024x1024xf32 || FAIL=1
36+
37+
# element wise
38+
python3 -m benchgc --verbose 0 --driver linalg --case abs --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
39+
python3 -m benchgc --verbose 0 --driver linalg --case ceil --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
40+
python3 -m benchgc --verbose 0 --driver linalg --case erf --md 0:1024x512xf32 --md 1:1024x512xf32 || FAIL=1
41+
python3 -m benchgc --verbose 0 --driver linalg --case floor --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
42+
python3 -m benchgc --verbose 0 --driver linalg --case log --md 0:4096x32xf32 --md 1:4096x32xf32 || FAIL=1
43+
python3 -m benchgc --verbose 0 --driver linalg --case negf --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
44+
python3 -m benchgc --verbose 0 --driver linalg --case exp --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
45+
python3 -m benchgc --verbose 0 --driver linalg --case round --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
46+
# python3 -m benchgc --verbose 0 --driver linalg --case rsqrt --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
47+
python3 -m benchgc --verbose 0 --driver linalg --case sqrt --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
48+
python3 -m benchgc --verbose 0 --driver linalg --case square --md 0:32x4096xf32 --md 1:32x4096xf32 || FAIL=1
49+
python3 -m benchgc --verbose 0 --driver linalg --case tanh --md 0:128x128xf32 --md 1:128x128xf32 || FAIL=1
50+
51+
# conv
52+
python3 -m benchgc --verbose 0 --driver linalg --case conv_1d_ncw_fcw --md 0:4x4x32xf32 --md 1:8x4x4xf32 --md 2:4x8x13xf32 --strides 2 --dilations 2 || FAIL=1
53+
python3 -m benchgc --verbose 0 --driver linalg --case conv_1d_nwc_wcf --md 0:4x32x4xf32 --md 1:4x4x8xf32 --md 2:4x13x8xf32 --strides 2 --dilations 2 || FAIL=1
54+
python3 -m benchgc --verbose 0 --driver linalg --case conv_1d --md 0:32xf32 --md 1:4xf32 --md 2:29xf32 || FAIL=1
55+
python3 -m benchgc --verbose 0 --driver linalg --case conv_2d_nchw_fchw --md 0:4x4x32x32xf32 --md 1:8x4x4x4xf32 --md 2:4x8x13x13xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
56+
python3 -m benchgc --verbose 0 --driver linalg --case conv_2d_ngchw_fgchw --md 0:4x2x2x32x32xf32 --md 1:4x2x2x4x4xf32 --md 2:4x2x4x13x13xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
57+
python3 -m benchgc --verbose 0 --driver linalg --case conv_2d_ngchw_gfchw --md 0:4x2x2x32x32xf32 --md 1:2x4x2x4x4xf32 --md 2:4x2x4x13x13xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
58+
python3 -m benchgc --verbose 0 --driver linalg --case conv_2d_nhwc_fhwc --md 0:4x32x32x4xf32 --md 1:8x4x4x4xf32 --md 2:4x13x13x8xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
59+
python3 -m benchgc --verbose 0 --driver linalg --case conv_2d_nhwc_hwcf --md 0:4x32x32x4xf32 --md 1:4x4x4x8xf32 --md 2:4x13x13x8xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
60+
python3 -m benchgc --verbose 0 --driver linalg --case conv_2d --md 0:32x32xf32 --md 1:4x4xf32 --md 2:29x29xf32 || FAIL=1
61+
python3 -m benchgc --verbose 0 --driver linalg --case conv_3d_ncdhw_fcdhw --md 0:4x4x32x32x32xf32 --md 1:8x4x4x4x4xf32 --md 2:4x8x13x13x13xf32 --strides 2 --strides 2 --strides 2 --dilations 2 --dilations 2 --dilations 2 || FAIL=1
62+
python3 -m benchgc --verbose 0 --driver linalg --case conv_3d_ndhwc_dhwcf --md 0:4x32x32x32x4xf32 --md 1:4x4x4x4x8xf32 --md 2:4x13x13x13x8xf32 --strides 2 --strides 2 --strides 2 --dilations 2 --dilations 2 --dilations 2 || FAIL=1
63+
python3 -m benchgc --verbose 0 --driver linalg --case conv_3d --md 0:32x32x32xf32 --md 1:4x4x4xf32 --md 2:29x29x29xf32 || FAIL=1
64+
65+
# depthwise conv
66+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_1d_ncw_cw --md 0:4x4x32xf32 --md 1:4x4xf32 --md 2:4x4x13xf32 --strides 2 --dilations 2 || FAIL=1
67+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_1d_nwc_wc --md 0:4x32x4xf32 --md 1:4x4xf32 --md 2:4x13x4xf32 --strides 2 --dilations 2 || FAIL=1
68+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_1d_nwc_wcm --md 0:4x32x4xf32 --md 1:4x4x3xf32 --md 2:4x13x4x3xf32 --strides 2 --dilations 2 || FAIL=1
69+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_2d_nchw_chw --md 0:4x4x32x32xf32 --md 1:4x4x4xf32 --md 2:4x4x13x13xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
70+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_2d_nhwc_hwc --md 0:4x32x32x4xf32 --md 1:4x4x4xf32 --md 2:4x13x13x4xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
71+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_2d_nhwc_hwcm --md 0:4x32x32x4xf32 --md 1:4x4x4x3xf32 --md 2:4x13x13x4x3xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
72+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_3d_ncdhw_cdhw --md 0:4x4x32x32x32xf32 --md 1:4x4x4x4xf32 --md 2:4x4x13x13x13xf32 --strides 2 --strides 2 --strides 2 --dilations 2 --dilations 2 --dilations 2 || FAIL=1
73+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_3d_ndhwc_dhwc --md 0:4x32x32x32x4xf32 --md 1:4x4x4x4xf32 --md 2:4x13x13x13x4xf32 --strides 2 --strides 2 --strides 2 --dilations 2 --dilations 2 --dilations 2 || FAIL=1
74+
python3 -m benchgc --verbose 0 --driver linalg --case depthwise_conv_3d_ndhwc_dhwcm --md 0:4x32x32x32x4xf32 --md 1:4x4x4x4x3xf32 --md 2:4x13x13x13x4x3xf32 --strides 2 --strides 2 --strides 2 --dilations 2 --dilations 2 --dilations 2 || FAIL=1
75+
76+
# pool
77+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_nchw_max --md 0:4x4x32x32xf32 --md 1:4x4xf32 --md 2:4x4x13x13xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
78+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_nchw_sum --md 0:4x4x32x32xf32 --md 1:4x4xf32 --md 2:4x4x13x13xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
79+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_ncw_max --md 0:4x4x32xf32 --md 1:4xf32 --md 2:4x4x13xf32 --strides 2 --dilations 2 || FAIL=1
80+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_ncw_sum --md 0:4x4x32xf32 --md 1:4xf32 --md 2:4x4x13xf32 --strides 2 --dilations 2 || FAIL=1
81+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_ndhwc_max --md 0:4x32x32x32x4xf32 --md 1:4x4x4xf32 --md 2:4x13x13x13x4xf32 --strides 2 --strides 2 --strides 2 --dilations 2 --dilations 2 --dilations 2 || FAIL=1
82+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_ndhwc_sum --md 0:4x32x32x32x4xf32 --md 1:4x4x4xf32 --md 2:4x13x13x13x4xf32 --strides 2 --strides 2 --strides 2 --dilations 2 --dilations 2 --dilations 2 || FAIL=1
83+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_nhwc_max --md 0:4x32x32x4xf32 --md 1:4x4xf32 --md 2:4x13x13x4xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
84+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_nhwc_sum --md 0:4x32x32x4xf32 --md 1:4x4xf32 --md 2:4x13x13x4xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
85+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_nhwc_min --md 0:4x32x32x4xf32 --md 1:4x4xf32 --md 2:4x13x13x4xf32 --strides 2 --strides 2 --dilations 2 --dilations 2 || FAIL=1
86+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_nwc_max --md 0:4x32x4xf32 --md 1:4xf32 --md 2:4x13x4xf32 --strides 2 --dilations 2 || FAIL=1
87+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_nwc_sum --md 0:4x32x4xf32 --md 1:4xf32 --md 2:4x13x4xf32 --strides 2 --dilations 2 || FAIL=1
88+
python3 -m benchgc --verbose 0 --driver linalg --case pooling_nwc_min --md 0:4x32x4xf32 --md 1:4xf32 --md 2:4x13x4xf32 --strides 2 --dilations 2 || FAIL=1
89+
90+
# softmax
91+
# python3 -m benchgc --verbose 0 --driver linalg --case softmax --md 0:32x4096xf32 --md 1:32x4096xf32 --dimension 1 || FAIL=1
92+
93+
# mlir
94+
# python3 -m benchgc --verbose 0 --driver mlir --case ${CASE_DIR}/llama2.mlir || FAIL=1
95+
96+
set +e
97+
exit $FAIL

‎test/CMakeLists.txt

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -6,3 +6,7 @@ endif ()
66
include(gtest)
77
add_subdirectory(dnnl)
88
add_subdirectory(mlir)
9+
10+
if(GC_BENCH_ENABLE)
11+
add_subdirectory(benchgc)
12+
endif()

‎test/benchgc/.gitignore

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,5 @@
1+
dist/
2+
src/benchgc.egg-info/
3+
build
4+
benchgc.egg-info/
5+
__pycache__

‎test/benchgc/CMakeLists.txt

Lines changed: 41 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,41 @@
1+
################################################################################
2+
# Copyright (C) 2024 Intel Corporation
3+
#
4+
# Licensed under the Apache License, Version 2.0 (the "License");
5+
# you may not use this file except in compliance with the License.
6+
# You may obtain a copy of the License at
7+
#
8+
# http://www.apache.org/licenses/LICENSE-2.0
9+
#
10+
# Unless required by applicable law or agreed to in writing,
11+
# software distributed under the License is distributed on an "AS IS" BASIS,
12+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13+
# See the License for the specific language governing permissions
14+
# and limitations under the License.
15+
# SPDX-License-Identifier: Apache-2.0
16+
################################################################################
17+
18+
if(NOT GC_BENCH_ENABLE)
19+
message(STATUS "Benchgc is not enabled")
20+
return()
21+
endif()
22+
23+
configure_file(setup.py ${CMAKE_BINARY_DIR}/test/benchgc/setup.py COPYONLY)
24+
25+
file(GLOB PYTHON_SCRIPTS "*.py")
26+
foreach(PY_SCRIPT ${PYTHON_SCRIPTS})
27+
configure_file(${PY_SCRIPT} ${CMAKE_BINARY_DIR} COPYONLY)
28+
endforeach()
29+
30+
find_package(Python 3.10 REQUIRED COMPONENTS Interpreter NumPy REQUIRED)
31+
add_custom_target(benchgc
32+
COMMAND ${Python_EXECUTABLE} setup.py bdist_wheel
33+
WORKING_DIRECTORY "${CMAKE_BINARY_DIR}/test/benchgc/"
34+
DEPENDS GcPythonModules)
35+
36+
add_subdirectory("src/benchgc")
37+
add_subdirectory("src/benchgc/arg")
38+
add_subdirectory("src/benchgc/mlir")
39+
add_subdirectory("src/benchgc/linalg")
40+
add_subdirectory("src/benchgc/tensor")
41+
add_subdirectory("src/benchgc/arith")

‎test/benchgc/README.md

Lines changed: 103 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,103 @@
1+
# benchgc - benchmark tool for graph compiler
2+
3+
## Description
4+
5+
Benchgc is a tool used to verify the correctness and performance of graph compiler. Benchgc accepts MLIR files based on the OneDNN graph dialect as test cases and prepares test data for them. For correctness verification, Benchgc will use PyTorch as a reference for comparison.
6+
7+
## Prerequisite
8+
* python >= 3.10
9+
* torch >= 2.2
10+
* pybind11
11+
12+
## Build and install
13+
```
14+
# Please execute at the top level of the project
15+
16+
mkdir -p build
17+
cd build
18+
19+
cmake .. -DMLIR_DIR=$MLIR_PATH -DGC_TEST_ENABLE=ON -DGC_ENABLE_BINDINGS_PYTHON=ON -DGC_BENCH_ENABLE=ON
20+
make -j benchgc
21+
22+
python -m pip install test/benchgc/dist/benchgc-*.whl
23+
24+
```
25+
26+
## Synopsis
27+
```
28+
python -m benchgc [OPTIONS] --driver [DRIVER] --case [CASE]
29+
```
30+
## Flags
31+
### --driver [str]
32+
* linalg: test the single op in linalg dialect
33+
* mlir: upload a mlir file and run
34+
* pattern: predefined pattern test such as mlp
35+
36+
### --case [str]
37+
* if driver=mlir, please provide a mlir file here to test
38+
* if driver=pattern, please provide the pre-defined pattern name, such as mlp here
39+
* if driver is a dialect name, please provide the detail op name to start a single op test
40+
41+
### --seed [int]
42+
* set the seed to generate the test data and reprodce the test
43+
44+
### --verbose [int]
45+
* set the verbose level
46+
47+
### --md index:SHAPExTYPE
48+
* Describe the shape and data type for argument
49+
* Not available when driver=mlir
50+
* index means the order of argument, including both inputs and outs
51+
* use prefix `0x` (e.g. `0xbf16`) to represent 0d memref or tensor input
52+
* use data type directly (e.g.`f32`) to represent a normal scalar
53+
54+
```
55+
# %arg0 -> index = 0
56+
# tensor<2x2x2xf32> -> index = 1
57+
58+
module {
59+
func.func @entry(%arg0: f32) -> tensor<2x2x2xf32> attributes {llvm.emit_c_interface} {
60+
%0 = tensor.empty() : tensor<2x2x2xf32>
61+
%1 = linalg.fill ins(%arg0 : f32) outs(%0 : tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
62+
return %1 : tensor<2x2x2xf32>
63+
}
64+
}
65+
```
66+
67+
### --fill index:fill_type:[:fill_parameter]*
68+
* If not set, benchgc will assign a default method for the argument
69+
70+
| description | fill_type | fill_parameter |
71+
|-------------|-----------|-----------|
72+
| Zero | Z | |
73+
| Normal | N | mean, std |
74+
| Poisson | P | lambda |
75+
| Binomial | B | n, p |
76+
| Uniform | U | a, b |
77+
| Integer | I | a, b |
78+
| Pytorch tensor dump | F | dump filename |
79+
| Benchdnn driver | D | driver_name[:driver filling parameter]* |
80+
81+
#### Benchdnn driver filling
82+
83+
| driver_name | driver filling parameter |
84+
|-------------|--------------------------|
85+
| binary | src0/src1:src0 dtype:src1 dtype:dst dtype |
86+
| conv | src/wei:src dtype:wei dtype:dst dtype:amplifier |
87+
| eltwise | algorithm: alpha: beta (please check https://oneapi-src.github.io/oneDNN/dev_guide_eltwise.html) |
88+
| matmul | src/wei:src dtype:wei dtype:dst dtype:amplifier |
89+
| pool | not required |
90+
91+
### --cmp index:cmp_type:[:cmp_parameter]*
92+
* If not set, benchgc will assign a default method for the argument
93+
94+
| description | cmp_type | cmp_parameter |
95+
|-------------|-----------|-----------|
96+
| P2P check | P | threshold, zero_percent(mistrust check) |
97+
| Norm check | N | threshold |
98+
| Benchdnn driver | D | driver_name:dtype:case |
99+
100+
## Example
101+
```
102+
python3 -m benchgc --verbose 0 --driver linalg --case add --md 0:4x5x6xf32 --md 1:4x5x6xf32 --md 2:4x5x6xf32
103+
```

0 commit comments

Comments
 (0)