Open
feat(qbmmv3): ScaleNZ layout support for MX ScaleB #1
HuangRZzzzAdd on Aug 19
feat(qbmmv3): ScaleNZ layout support for MX ScaleB #1
Open
HuangRZzzzAdd on Aug 19
Merged fromscaleB-nz-layoutintomaster
Total: 771 files changed+49613-19895
@@ -123,34 +123,47 @@ case "${task_name}" in
123 touch build_out/cann-ops-nn-950_linux-aarch64.run123 touch build_out/cann-ops-nn-950_linux-aarch64.run
124 fi124 fi
125 ;;125 ;;
126-esac126+ Compile_Ascend_X86_mobile_station)
127- 127+ if [ "${TARGET_BRANCH}" = "master" ];then
128-case "${task_name}|${TARGET_BRANCH}" in128+ wget -nv https://kiri-obs.obs.cn-north-4.myhuaweicloud.com/Cann%20Large%20Model%20Foundation%208.5.0.rc002/cann-bisheng-compiler_9.2.0_linux-x86_64.run
129- "Compile_Ascend_X86_mobile_station|master")129+ chmod +x *.run
130- wget -nv https://kiri-obs.obs.cn-north-4.myhuaweicloud.com/Cann%20Large%20Model%20Foundation%208.5.0.rc002/cann-bisheng-compiler_9.2.0_linux-x86_64.run130+ sudo -u jenkins ./*.run --full --quiet --install-path=/home/jenkins/Ascend
131- chmod +x *.run131+ bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16
132- sudo -u jenkins ./*.run --full --quiet --install-path=/home/jenkins/Ascend132+ echo "exec cmd: [bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16]"
133- bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16133+ else
134- echo "exec cmd: [bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16]"134+ echo "not need build mobile_station"
135+ mkdir build_out
136+ touch build_out/cann-ops-nn-kirinx90_linux-x86_64.run
137+ exit 0
138+ fi
135 ;;139 ;;
136- *)140+ Compile_Ascend_X86_mobile_station_ubuntu24)
137- echo "not need build single"141+ if [ "${TARGET_BRANCH}" = "master" ];then
138- mkdir single142+ wget -nv https://kiri-obs.obs.cn-north-4.myhuaweicloud.com/Cann%20Large%20Model%20Foundation%208.5.0.rc002/cann-bisheng-compiler_9.2.0_linux-x86_64.run
139- touch single.tar.gz143+ chmod +x *.run
144+ sudo -u jenkins ./*.run --full --quiet --install-path=/home/jenkins/Ascend
145+ bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16
146+ echo "exec cmd: [bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16]"
147+ else
148+ echo "not need build mobile_station"
149+ mkdir build_out
150+ touch build_out/cann-ops-nn-kirinx90_linux-x86_64.run
151+ exit 0
152+ fi
140 ;;153 ;;
141-esac154+ Compile_Ascend_X86_mobile_station_9030_ubuntu24)
142-case "${task_name}|${TARGET_BRANCH}" in155+ if [ "${TARGET_BRANCH}" = "master" ];then
143- "Compile_Ascend_X86_mobile_station_ubuntu24|master")156+ wget -nv https://kiri-obs.obs.cn-north-4.myhuaweicloud.com/Cann%20Large%20Model%20Foundation%208.5.0.rc002/cann-bisheng-compiler_9.2.0_linux-x86_64.run
144- wget -nv https://kiri-obs.obs.cn-north-4.myhuaweicloud.com/Cann%20Large%20Model%20Foundation%208.5.0.rc002/cann-bisheng-compiler_9.2.0_linux-x86_64.run157+ chmod +x *.run
145- chmod +x *.run158+ sudo -u jenkins ./*.run --full --quiet --install-path=/home/jenkins/Ascend
146- sudo -u jenkins ./*.run --full --quiet --install-path=/home/jenkins/Ascend159+ bash build.sh --pkg --soc=kirinx9030 --cann_3rd_lib_path=/home/jenkins/opensource -j16
147- bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16160+ echo "exec cmd: [bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16]"
148- echo "exec cmd: [bash build.sh --pkg --soc=kirinx90 --cann_3rd_lib_path=/home/jenkins/opensource -j16]"161+ else
149- ;;162+ echo "not need build mobile_station"
150- *)163+ mkdir build_out
151- echo "not need build single"164+ touch build_out/cann-ops-nn-kirinx90_linux-x86_64.run
152- mkdir single165+ exit 0
153- touch single.tar.gz166+ fi
154 ;;167 ;;
155esac168esac
156 169 
@@ -12,8 +12,7 @@
12set -euo pipefail12set -euo pipefail
13 13 
14echo "start run test case, please wait ..."14echo "start run test case, please wait ..."
15-cd /home/taskspace15+ 
16-WORKSPACE=/home/taskspace
17 16 
18export ASCEND_GLOBAL_LOG_LEVEL=217export ASCEND_GLOBAL_LOG_LEVEL=2
19export ASCEND_SLOG_PRINT_TO_STDOUT=018export ASCEND_SLOG_PRINT_TO_STDOUT=0
@@ -49,7 +49,7 @@ jobs:
49 artifact-path: "${{ steps.process_checkout.outputs.path }}/pre-commit*.txt"49 artifact-path: "${{ steps.process_checkout.outputs.path }}/pre-commit*.txt"
50 object-prefix: ${{ env.obs_path }}/50 object-prefix: ${{ env.obs_path }}/
51 if: "${{ default() }}"51 if: "${{ default() }}"
52- runs-on: [dedicate-hosted, x64, small]52+ runs-on: [self-hosted, arch=x64, spec=small]
53 container:53 container:
54 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.precommit_image_version }}54 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.precommit_image_version }}
55 JOB_check_pr:55 JOB_check_pr:
@@ -79,7 +79,7 @@ jobs:
79 with:79 with:
80 repo_name: ${{ env.repo_name }}80 repo_name: ${{ env.repo_name }}
81 workspace: ${{ steps.process_checkout.outputs.path }}81 workspace: ${{ steps.process_checkout.outputs.path }}
82- runs-on: [dedicate-hosted, x64, small]82+ runs-on: [self-hosted, arch=x64, spec=small]
83 container:83 container:
84 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.codecheck_image_version }}84 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.codecheck_image_version }}
85 JOB_sca:85 JOB_sca:
@@ -107,7 +107,7 @@ jobs:
107 cd CI/scripts107 cd CI/scripts
108 python3 codescan_gitcode.py --access_key ${{secrets.SCA_AK}} --secret_access_key ${{secrets.SCA_SK}} --access_key_get ${{secrets.SCA_AK_GET}} --secret_access_key_get ${{secrets.SCA_SK_GET}} --org_name ${org_name} --pr_id ${MERGE_ID} --repo ${repo_name} --commit_id ${{ atomgit.event.pull_request.merge_commit_sha }}108 python3 codescan_gitcode.py --access_key ${{secrets.SCA_AK}} --secret_access_key ${{secrets.SCA_SK}} --access_key_get ${{secrets.SCA_AK_GET}} --secret_access_key_get ${{secrets.SCA_SK_GET}} --org_name ${org_name} --pr_id ${MERGE_ID} --repo ${repo_name} --commit_id ${{ atomgit.event.pull_request.merge_commit_sha }}
109 if: "${{ default() }}"109 if: "${{ default() }}"
110- runs-on: [dedicate-hosted, x64, small]110+ runs-on: [self-hosted, arch=x64, spec=small]
111 container:111 container:
112 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.codecheck_image_version }}112 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.codecheck_image_version }}
113 JOB_antiposion:113 JOB_antiposion:
@@ -135,7 +135,7 @@ jobs:
135 cd CI/scripts135 cd CI/scripts
136 python3 anti_virus_gitcode.py --access_key ${{secrets.ANTI_AK}} --secret_access_key ${{secrets.ANTI_SK}} --access_key_get ${{secrets.ANTI_AK_GET}} --secret_access_key_get ${{secrets.ANTI_SK_GET}} --org_name ${org_name} --pr_id ${MERGE_ID} --repo ${repo_name} --commit_id ${{ atomgit.event.pull_request.merge_commit_sha }}136 python3 anti_virus_gitcode.py --access_key ${{secrets.ANTI_AK}} --secret_access_key ${{secrets.ANTI_SK}} --access_key_get ${{secrets.ANTI_AK_GET}} --secret_access_key_get ${{secrets.ANTI_SK_GET}} --org_name ${org_name} --pr_id ${MERGE_ID} --repo ${repo_name} --commit_id ${{ atomgit.event.pull_request.merge_commit_sha }}
137 if: "${{ default() }}"137 if: "${{ default() }}"
138- runs-on: [dedicate-hosted, x64, small]138+ runs-on: [self-hosted, arch=x64, spec=small]
139 container:139 container:
140 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.codecheck_image_version }}140 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.codecheck_image_version }}
141 JOB_codecheck:141 JOB_codecheck:
@@ -163,6 +163,6 @@ jobs:
163 cd CI/scripts163 cd CI/scripts
164 python3 codecheck_gitcode.py --access_key ${{secrets.CODECHECK_AK}} --secret_access_key ${{secrets.CODECHECK_SK}} --access_key_get ${{secrets.CODECHECK_AK_GET}} --secret_access_key_get ${{secrets.CODECHECK_SK_GET}} --org_name ${org_name} --pr_id ${MERGE_ID} --repo ${repo_name} --commit_id ${{ atomgit.event.pull_request.merge_commit_sha }}164 python3 codecheck_gitcode.py --access_key ${{secrets.CODECHECK_AK}} --secret_access_key ${{secrets.CODECHECK_SK}} --access_key_get ${{secrets.CODECHECK_AK_GET}} --secret_access_key_get ${{secrets.CODECHECK_SK_GET}} --org_name ${org_name} --pr_id ${MERGE_ID} --repo ${repo_name} --commit_id ${{ atomgit.event.pull_request.merge_commit_sha }}
165 if: "${{ default() }}"165 if: "${{ default() }}"
166- runs-on: [dedicate-hosted, x64, small]166+ runs-on: [self-hosted, arch=x64, spec=small]
167 container:167 container:
168 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.codecheck_image_version }}168 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.codecheck_image_version }}
@@ -66,14 +66,20 @@ jobs:
66 name: rename_package66 name: rename_package
67 if: ${{ inputs.package_name != '' }}67 if: ${{ inputs.package_name != '' }}
68 run: |68 run: |
69- cd ${{ steps.process_checkout.outputs.path }}/build_out69+ mkdir -p "${{ steps.process_checkout.outputs.path }}/build_out"
70- for f in *.run; do70+ cd "${{ steps.process_checkout.outputs.path }}/build_out"
71- [ -f "$f" ] || continue71+ if [[ "${{ inputs.task_name }}" == compile_single ]]; then
72- [ "$f" = "${{ inputs.package_name }}" ] && continue72+ mv ../single.tar.gz ${{ inputs.package_name }}
73- mv "$f" "${{ inputs.package_name }}"73+ echo "mv: ../single.tar.gz -> ${{ inputs.package_name }}"
74- echo "Renamed: $f -> ${{ inputs.package_name }}"74+ else
75- break75+ for f in *.run; do
76- done76+ [ -f "$f" ] || continue
77+ [ "$f" = "${{ inputs.package_name }}" ] && continue
78+ mv "$f" "${{ inputs.package_name }}"
79+ echo "Renamed: $f -> ${{ inputs.package_name }}"
80+ break
81+ done
82+ fi
77 -83 -
78 name: api-check84 name: api-check
79 identifier: api-check85 identifier: api-check
@@ -103,8 +109,8 @@ jobs:
103 object-prefix: ${{ env.obs_path }}/109 object-prefix: ${{ env.obs_path }}/
104 if: "${{ default() }}"110 if: "${{ default() }}"
105 runs-on:111 runs-on:
106- - dedicate-hosted112+ - self-hosted
107- - ${{ inputs.arch }}113+ - arch=${{ inputs.arch }}
108- - xlarge114+ - spec=xlarge
109 container:115 container:
110 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.image_version }}116 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.image_version }}
@@ -18,12 +18,12 @@ on:
18 keyword: '^(?:\/)?compile*'18 keyword: '^(?:\/)?compile*'
19 19 
20env:20env:
21- org_name: "cann"21+ org_name: "${{ vars.ORG_NAME }}"
22- repo_name: "ops-nn"22+ repo_name: ${{atomgit.event.repository.name}}
23 MERGE_ID: ${{atomgit.event.pull_request.number}}23 MERGE_ID: ${{atomgit.event.pull_request.number}}
24+ obs_path: "${{ env.org_name }}/${{ env.repo_name }}/ci/package/${{ env.MERGE_ID }}"
24 TARGET_BRANCH: ${{atomgit.ref_name}}25 TARGET_BRANCH: ${{atomgit.ref_name}}
25 SOURCE_BRANCH: ${{atomgit.head_ref}}26 SOURCE_BRANCH: ${{atomgit.head_ref}}
26- obs_path: "${{ vars.OBS_PATH }}/${{ env.MERGE_ID }}"
27 obs_smoke_path: "${{ env.obs_path }}/presmoke/2"27 obs_smoke_path: "${{ env.obs_path }}/presmoke/2"
28 online: ${{ vars.APPLY_LABEL }}28 online: ${{ vars.APPLY_LABEL }}
29 29 
@@ -46,7 +46,7 @@ stages:
46 run: |46 run: |
47 echo "image: ${{ steps.revise_image.outputs.Compile_Ascend_X86 }}"47 echo "image: ${{ steps.revise_image.outputs.Compile_Ascend_X86 }}"
48 if: "${{ default() }}"48 if: "${{ default() }}"
49- runs-on: ["codearts-hosted", "ubuntu-latest", "x64", "small"]49+ runs-on: ["self-hosted", "arch=x64", "spec=small"]
50 JOB_pr:50 JOB_pr:
51 name: Test_get-pr51 name: Test_get-pr
52 steps:52 steps:
@@ -123,7 +123,7 @@ stages:
123 - name: associate-pr-comment123 - name: associate-pr-comment
124 uses: associate-pr-comment124 uses: associate-pr-comment
125 if: env.online == 'true'125 if: env.online == 'true'
126- runs-on: [dedicate-hosted, x64, small]126+ runs-on: [self-hosted, arch=x64, spec=small]
127 container:127 container:
128 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/ubuntu20.04.05_x86:k8s_v1.1006128 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/ubuntu20.04.05_x86:k8s_v1.1006
129 pre:129 pre:
@@ -337,15 +337,6 @@ stages:
337 uses: checkout337 uses: checkout
338 with:338 with:
339 token: ${{secrets.GIT_TOKEN}}339 token: ${{secrets.GIT_TOKEN}}
340- path: "./runner"
341- -
342- name: Checkout gitcode
343- identifier: process_checkout_gitcode
344- uses: checkout
345- with:
346- repository: "https://gitcode.com/cann/.gitcode.git"
347- ref: master
348- path: "./runner/ci"
349 -340 -
350 name: download341 name: download
351 uses: obs-download342 uses: obs-download
@@ -360,37 +351,13 @@ stages:
360 ${{ env.obs_path }}/update_file_detail.txt351 ${{ env.obs_path }}/update_file_detail.txt
361 ${{ env.obs_path }}/pre_smoke.sh352 ${{ env.obs_path }}/pre_smoke.sh
362 path: ${{ steps.process_checkout.outputs.path }}353 path: ${{ steps.process_checkout.outputs.path }}
363- -
364- name: tar businesscode
365- run: |
366- echo "############${{ steps.process_checkout.outputs.path }}#################"
367- echo "#####image: swr.cn-south-1.myhuaweicloud.com/pre_smoke/${{ jobs.JOB_image.outputs.PreSmoke_A900 }}#####"
368- cd ${{ steps.process_checkout.outputs.path }}
369- tar -zcf ../smoke_data.tar.gz .
370- mv ../smoke_data.tar.gz ./
371- -
372- name: upload
373- identifier: process_upload
374- uses: obs-upload
375- with:
376- endpoint: "https://obs.cn-north-4.myhuaweicloud.com"
377- bucket: "ascend-ci"
378- access-key: ${{secrets.AK}}
379- secret-key: ${{secrets.SK}}
380- artifact-path: |
381- ${{ steps.process_checkout.outputs.path }}/smoke_data.tar.gz
382- object-prefix: ${{ env.obs_smoke_path }}/
383 -354 -
384 name: smoke executor355 name: smoke executor
385 timeout-minutes: 30356 timeout-minutes: 30
386- uses: manifest-management-plugin357+ run: |
387- with:
388- action: DEPLOY
389- repo: ${{ steps.process_checkout.outputs.path }}
390- file_path: ${{ steps.process_checkout.outputs.path }}/ci/smoke-conf/smoke-A2.yaml
391- manifest_image: swr.cn-south-1.myhuaweicloud.com/pre_smoke/${{ jobs.JOB_image.outputs.PreSmoke_A900 }}
392- custom_shell_script: |-
393 set -e358 set -e
359+ # set env
360+ export WORKSPACE=${{ steps.process_checkout.outputs.path }}
394 export single_tar_url=${{jobs.JOB_compile_single.outputs.arm_run_url}}361 export single_tar_url=${{jobs.JOB_compile_single.outputs.arm_run_url}}
395 export experimental_run_url=${{jobs.JOB_Compile_Ascend_experimental_ubuntu24.outputs.arm_run_url}}362 export experimental_run_url=${{jobs.JOB_Compile_Ascend_experimental_ubuntu24.outputs.arm_run_url}}
396 export fatrelu_run_url=${{jobs.JOB_Pre_compile.outputs.arm_run_url}}363 export fatrelu_run_url=${{jobs.JOB_Pre_compile.outputs.arm_run_url}}
@@ -398,21 +365,20 @@ stages:
398 export pr_id=${{ env.MERGE_ID }}365 export pr_id=${{ env.MERGE_ID }}
399 export AK=${{secrets.AK}}366 export AK=${{secrets.AK}}
400 export SK=${{secrets.SK}}367 export SK=${{secrets.SK}}
401- mkdir -p /home/taskspace && cd /home/taskspace368+ bash ${{ steps.process_checkout.outputs.path }}/pre_smoke.sh
402- wget -nv ${{ steps.process_upload.outputs.primary-url }}369+ - name: upload
403- tar -zxf $(basename "${{ steps.process_upload.outputs.primary-url }}")370+ uses: obs-upload
404- bash /home/taskspace/pre_smoke.sh
405- -
406- name: smoke resource clean
407- uses: manifest-management-plugin
408- if: ${{ always() }}
409 with:371 with:
410- action: CRUSH372+ endpoint: "https://obs.cn-north-4.myhuaweicloud.com"
411- repo: ${{ steps.process_checkout.outputs.path }}373+ bucket: "ascend-ci"
412- file_path: ${{ steps.process_checkout.outputs.path }}/ci/smoke-conf/smoke-A2.yaml374+ access-key: ${{ secrets.AK }}
413- 375+ secret-key: ${{ secrets.SK }}
414- if: env.TARGET_BRANCH == 'master'376+ artifact-path: "${{ steps.process_checkout.outputs.path }}/slog.tar.gz"
415- runs-on: ['self-hosted','arch=arm','smoke=204']377+ object-prefix: ${obs_smoke_path}/plog/
378+ if: "${{ default() }}"
379+ runs-on: ['self-hosted','npu=1','smoke=204']
380+ container:
381+ image: swr.cn-south-1.myhuaweicloud.com/pre_smoke/${{ jobs.JOB_image.outputs.PreSmoke_A900 }}
416 pre:382 pre:
417 - type: auto383 - type: auto
418 fail-fast: true384 fail-fast: true
@@ -0,0 +1,93 @@
1+name: st_test_950_action
2+ 
3+on:
4+ pr_comment:
5+ types: [created]
6+ keyword: '^(?:\/)?st_950\b'
7+ pull_request_comment:
8+ types: [created]
9+ branches: ['*']
10+ comments: ['^(?:\/)?st_950\b']
11+ workflow_call:
12+ inputs:
13+ x86_run_key:
14+ type: string
15+ required: false
16+ 
17+env:
18+ org_name: ${{ vars.ORG_NAME }}
19+ repo_name: "ops-nn"
20+ MERGE_ID: ${{atomgit.event.pull_request.number}}
21+ TARGET_BRANCH: ${{atomgit.ref_name}}
22+ SOURCE_BRANCH: ${{atomgit.head_ref}}
23+ obs_path: "${{ env.org_name }}/${{ env.repo_name }}/ci/package/${{ env.MERGE_ID }}"
24+ run_package_name: "cann-ops-nn-custom_linux-x86_64_950.run"
25+ 
26+jobs:
27+ JOB_st_test_950:
28+ name: ST_test_950
29+ steps:
30+ -
31+ name: Checkout
32+ identifier: process_checkout
33+ uses: checkout
34+ with:
35+ ref: ${{ atomgit.event.pull_request.merge_commit_sha || atomgit.sha }}
36+ token: ${{secrets.GIT_TOKEN}}
37+ -
38+ name: download_pr_filelist
39+ uses: obs-download
40+ with:
41+ endpoint: "https://obs.cn-north-4.myhuaweicloud.com"
42+ bucket: "ascend-ci"
43+ access-key: ${{ secrets.AK }}
44+ secret-key: ${{ secrets.SK }}
45+ key: |
46+ ${{ env.obs_path }}/pr_filelist.txt
47+ path: ${{ steps.process_checkout.outputs.path }}
48+ -
49+ name: download_compiled_package
50+ uses: obs-download
51+ with:
52+ endpoint: "https://obs.cn-north-4.myhuaweicloud.com"
53+ bucket: "ascend-ci"
54+ access-key: ${{ secrets.AK }}
55+ secret-key: ${{ secrets.SK }}
56+ key: |
57+ ${{ env.obs_path }}/${{ env.run_package_name }}
58+ path: ${{ steps.process_checkout.outputs.path }}
59+ -
60+ name: install_run_package
61+ run: |
62+ cd ${{ steps.process_checkout.outputs.path }}
63+ if [[ -f "${{ env.run_package_name }}" ]]; then
64+ echo "Installing compiled package: ${{ env.run_package_name }}"
65+ chmod +x "${{ env.run_package_name }}"
66+ source /usr/local/Ascend/cann/bin/setenv.bash 2>/dev/null || true
67+ echo 'y' | bash "${{ env.run_package_name }}" --quiet
68+ else
69+ echo "ERROR: compiled package not found: ${{ env.run_package_name }}"
70+ exit 1
71+ fi
72+ -
73+ name: run_st_test
74+ identifier: st_test
75+ run: |
76+ export WORKSPACE=${{ steps.process_checkout.outputs.path }}
77+ cd ${{ steps.process_checkout.outputs.path }}
78+ source /usr/local/Ascend/cann/bin/setenv.bash 2>/dev/null || true
79+ export CI_ONLINE_ST=1
80+ bash scripts/ci/ops_st_test.sh --pr_filelist=./pr_filelist.txt --soc_version=ascend950
81+ -
82+ name: upload_results
83+ uses: obs-upload
84+ with:
85+ endpoint: "https://obs.cn-north-4.myhuaweicloud.com"
86+ bucket: "ascend-ci"
87+ access-key: ${{secrets.AK}}
88+ secret-key: ${{secrets.SK}}
89+ artifact-path: |
90+ ${{ steps.process_checkout.outputs.path }}/st/log/**/*
91+ object-prefix: ${{ env.obs_path }}/st_results/
92+ if: "${{ default() }}"
93+ runs-on: ['self-hosted', 'os=ubuntu', 'evb=a5', 'repo=nn']
@@ -61,6 +61,6 @@ jobs:
61 ${{ steps.process_checkout.outputs.path }}/StaticCheck_*.json61 ${{ steps.process_checkout.outputs.path }}/StaticCheck_*.json
62 object-prefix: ${{ env.obs_path }}/62 object-prefix: ${{ env.obs_path }}/
63 if: "${{ default() }}"63 if: "${{ default() }}"
64- runs-on: [dedicate-hosted, x64, small]64+ runs-on: [self-hosted, arch=x64, spec=small]
65 container:65 container:
66 image: swr.cn-north-4.myhuaweicloud.com/static_check/${{ inputs.image_version }}66 image: swr.cn-north-4.myhuaweicloud.com/static_check/${{ inputs.image_version }}
@@ -69,6 +69,6 @@ jobs:
69 ${{ steps.process_checkout.outputs.path }}/${{ inputs.ut_package }}69 ${{ steps.process_checkout.outputs.path }}/${{ inputs.ut_package }}
70 object-prefix: ${{ env.obs_path }}/70 object-prefix: ${{ env.obs_path }}/
71 if: "${{ default() }}"71 if: "${{ default() }}"
72- runs-on: [dedicate-hosted, x64, xlarge]72+ runs-on: [self-hosted, arch=x64, spec=xlarge]
73 container:73 container:
74 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.image_version }}74 image: swr.cn-north-4.myhuaweicloud.com/ci_cann/${{ inputs.image_version }}
@@ -123,6 +123,7 @@ if(DOWNLOAD_OPS_TEST_KIT)
123 include(cmake/third_party/ops_test_kit.cmake)123 include(cmake/third_party/ops_test_kit.cmake)
124endif()124endif()
125include(cmake/dependencies.cmake)125include(cmake/dependencies.cmake)
126+init_cann_project()
126include(cmake/third_party/ops-tensor.cmake)127include(cmake/third_party/ops-tensor.cmake)
127add_cann_third_party(json)128add_cann_third_party(json)
128add_cann_third_party(abseil-cpp)129add_cann_third_party(abseil-cpp)
@@ -139,7 +140,6 @@ if(ENABLE_CUSTOM AND ENABLE_ASC_BUILD)
139 include(cmake/custom_kernel.cmake)140 include(cmake/custom_kernel.cmake)
140endif()141endif()
141 142 
142-init_cann_project()
143if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)143if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
144 check_cann_pkg_build_deps("ops-nn")144 check_cann_pkg_build_deps("ops-nn")
145endif()145endif()
@@ -9,9 +9,13 @@
9# -----------------------------------------------------------------------------------------------------------9# -----------------------------------------------------------------------------------------------------------
10 10 
11__all__ = [11__all__ = [
12+ "situ_glu",
13+ "situ_glu_grad",
12 "swiglu_group",14 "swiglu_group",
13 "swiglu_group_quant",15 "swiglu_group_quant",
14]16]
15 17 
18+from .situ_glu import situ_glu
19+from .situ_glu_grad import situ_glu_grad
16from .swiglu_group import swiglu_group20from .swiglu_group import swiglu_group
17from .swiglu_group_quant import swiglu_group_quant21from .swiglu_group_quant import swiglu_group_quant
Rcommon/src/framework/celu_onnx_plugin.cpp→activation/celu_v2/framework/celu_onnx_plugin.cpp+44-44
@@ -1,44 +1,44 @@
1-/**1+/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3- * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4- * CANN Open Software License Agreement Version 2.0 (the "License").4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5- * Please refer to the License for details. You may not use this file except in compliance with the License.5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6- * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7- * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8- * See LICENSE in the root of the software repository for the full text of the License.8+ * See LICENSE in the root of the software repository for the full text of the License.
9- */9+ */
10- 10+ 
11-#include "onnx_common.h"11+#include "onnx_common.h"
12- 12+ 
13-namespace domi {13+namespace domi {
14-using NodeProto = ge::onnx::NodeProto;14+using NodeProto = ge::onnx::NodeProto;
15-static Status ParseParamsCelu(const Message* op_src, ge::Operator& op_dest)15+static Status ParseParamsCelu(const Message* op_src, ge::Operator& op_dest)
16-{16+{
17- const NodeProto* node = reinterpret_cast<const NodeProto*>(op_src);17+ const NodeProto* node = reinterpret_cast<const NodeProto*>(op_src);
18- if (node == nullptr) {18+ if (node == nullptr) {
19- OP_LOGE(GetOpName(op_dest).c_str(), "Dynamic cast op_src to NodeProto failed.");19+ OP_LOGE(GetOpName(op_dest).c_str(), "Dynamic cast op_src to NodeProto failed.");
20- return FAILED;20+ return FAILED;
21- }21+ }
22- float alpha_value = 1.0;22+ float alpha_value = 1.0;
23- op_dest.SetAttr("alpha", alpha_value);23+ op_dest.SetAttr("alpha", alpha_value);
24- for (const auto& attr : node->attribute()) {24+ for (const auto& attr : node->attribute()) {
25- if (attr.name() == "alpha" && attr.type() == ge::onnx::AttributeProto::FLOAT) {25+ if (attr.name() == "alpha" && attr.type() == ge::onnx::AttributeProto::FLOAT) {
26- alpha_value = attr.f();26+ alpha_value = attr.f();
27- op_dest.SetAttr("alpha1", alpha_value);27+ op_dest.SetAttr("alpha1", alpha_value);
28- op_dest.SetAttr("alpha2", alpha_value);28+ op_dest.SetAttr("alpha2", alpha_value);
29- }29+ }
30- }30+ }
31- return SUCCESS;31+ return SUCCESS;
32-}32+}
33-// register Celu op info to GE33+// register Celu op info to GE
34-REGISTER_CUSTOM_OP("Celu")34+REGISTER_CUSTOM_OP("Celu")
35- .FrameworkType(ONNX)35+ .FrameworkType(ONNX)
36- .OriginOpType({ge::AscendString("ai.onnx::8::Celu"), ge::AscendString("ai.onnx::9::Celu"),36+ .OriginOpType({ge::AscendString("ai.onnx::8::Celu"), ge::AscendString("ai.onnx::9::Celu"),
37- ge::AscendString("ai.onnx::10::Celu"), ge::AscendString("ai.onnx::11::Celu"),37+ ge::AscendString("ai.onnx::10::Celu"), ge::AscendString("ai.onnx::11::Celu"),
38- ge::AscendString("ai.onnx::12::Celu"), ge::AscendString("ai.onnx::13::Celu"),38+ ge::AscendString("ai.onnx::12::Celu"), ge::AscendString("ai.onnx::13::Celu"),
39- ge::AscendString("ai.onnx::14::Celu"), ge::AscendString("ai.onnx::15::Celu"),39+ ge::AscendString("ai.onnx::14::Celu"), ge::AscendString("ai.onnx::15::Celu"),
40- ge::AscendString("ai.onnx::16::Celu"), ge::AscendString("ai.onnx::17::Celu"),40+ ge::AscendString("ai.onnx::16::Celu"), ge::AscendString("ai.onnx::17::Celu"),
41- ge::AscendString("ai.onnx::18::Celu")})41+ ge::AscendString("ai.onnx::18::Celu")})
42- .ParseParamsFn(ParseParamsCelu)42+ .ParseParamsFn(ParseParamsCelu)
43- .ImplyType(ImplyType::TVM);43+ .ImplyType(ImplyType::TVM);
44-} // namespace domi44+} // namespace domi
@@ -60,18 +60,18 @@ __aicore__ inline void ReduceAllVf(LocalTensor<int64_t>& reduceSumUb, LocalTenso
60 AscendC::MicroAPI::MaskReg mask = AscendC::MicroAPI::CreateMask<int64_t, MicroAPI::MaskPattern::ALL>();60 AscendC::MicroAPI::MaskReg mask = AscendC::MicroAPI::CreateMask<int64_t, MicroAPI::MaskPattern::ALL>();
61 for (uint16_t i = 0; i < times; i++) {61 for (uint16_t i = 0; i < times; i++) {
62 AscendC::MicroAPI::AddrReg srcIdxOffset = AscendC::MicroAPI::CreateAddrReg<int64_t>(i, vfTidx);62 AscendC::MicroAPI::AddrReg srcIdxOffset = AscendC::MicroAPI::CreateAddrReg<int64_t>(i, vfTidx);
63- AscendC::MicroAPI::DataCopy(srcReg, srcAddr, srcIdxOffset);63+ AscendC::MicroAPI::LoadAlign(srcReg, srcAddr, srcIdxOffset);
64 AscendC::MicroAPI::Add(addReg, addReg, srcReg, mask);64 AscendC::MicroAPI::Add(addReg, addReg, srcReg, mask);
65 }65 }
66- AscendC::MicroAPI::ReduceSum(reduceSumReg, addReg, mask);66+ AscendC::MicroAPI::Reduce<ReduceType::SUM>(reduceSumReg, addReg, mask);
67 for (uint16_t j = 0; j < tailTimes; j++) {67 for (uint16_t j = 0; j < tailTimes; j++) {
68 AscendC::MicroAPI::MaskReg maskT = AscendC::MicroAPI::UpdateMask<int64_t>(tailNum);68 AscendC::MicroAPI::MaskReg maskT = AscendC::MicroAPI::UpdateMask<int64_t>(tailNum);
69- AscendC::MicroAPI::DataCopy(srcReg, srcAddr1);69+ AscendC::MicroAPI::LoadAlign(srcReg, srcAddr1);
70- AscendC::MicroAPI::ReduceSum(reduceSumTReg, srcReg, maskT);70+ AscendC::MicroAPI::Reduce<ReduceType::SUM>(reduceSumTReg, srcReg, maskT);
71 AscendC::MicroAPI::Add(reduceSumReg, reduceSumTReg, reduceSumReg, maskT);71 AscendC::MicroAPI::Add(reduceSumReg, reduceSumTReg, reduceSumReg, maskT);
72 }72 }
73 AscendC::MicroAPI::MaskReg maskOne = AscendC::MicroAPI::CreateMask<int64_t, MicroAPI::MaskPattern::VL1>();73 AscendC::MicroAPI::MaskReg maskOne = AscendC::MicroAPI::CreateMask<int64_t, MicroAPI::MaskPattern::VL1>();
74- AscendC::MicroAPI::DataCopy(dstAddr, reduceSumReg, maskOne);74+ AscendC::MicroAPI::StoreAlign(dstAddr, reduceSumReg, maskOne);
75 }75 }
76}76}
77 77 
@@ -85,11 +85,10 @@ public:
85 85 
86private:86private:
87 __aicore__ inline void ComputeTiling();87 __aicore__ inline void ComputeTiling();
88- __aicore__ inline void UbDataCopy(__local_mem__ T* inAddr, __local_mem__ T* outAddr, int64_t onceNum);88+ __aicore__ inline void UbStoreAlign(__ubuf__ T* inAddr, __ubuf__ T* outAddr, int64_t onceNum);
89 __aicore__ inline void CopyIn(int64_t gmOffset, int64_t count, int64_t blockLen);89 __aicore__ inline void CopyIn(int64_t gmOffset, int64_t count, int64_t blockLen);
90- __aicore__ inline void ComputeVfSwiglu(__local_mem__ T* x1UbAddr, __local_mem__ T* x2UbAddr,90+ __aicore__ inline void ComputeVfSwiglu(__ubuf__ T* x1UbAddr, __ubuf__ T* x2UbAddr, __ubuf__ T* swigluUbAddr,
91- __local_mem__ T* swigluUbAddr, int64_t dim0OnceSize, int64_t dim1OnceSize,91+ int64_t dim0OnceSize, int64_t dim1OnceSize, int64_t alignDim1In);
92- int64_t alignDim1In);
93 __aicore__ inline void CopyOut(int64_t gmOffset, int64_t count, int64_t blockLen);92 __aicore__ inline void CopyOut(int64_t gmOffset, int64_t count, int64_t blockLen);
94 93 
95private:94private:
@@ -238,7 +237,7 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::Process()
238 inQueX_.FreeTensor(inputUb);237 inQueX_.FreeTensor(inputUb);
239 LocalTensor<T> outUb = outQueY_.AllocTensor<T>();238 LocalTensor<T> outUb = outQueY_.AllocTensor<T>();
240 auto outUbAddr = (__ubuf__ T*)outUb.GetPhyAddr();239 auto outUbAddr = (__ubuf__ T*)outUb.GetPhyAddr();
241- UbDataCopy(swigluAddr, outUbAddr, onceNum);240+ UbStoreAlign(swigluAddr, outUbAddr, onceNum);
242 outQueY_.EnQue(outUb);241 outQueY_.EnQue(outUb);
243 CopyOut(outGmOffset, 1, onceNum);242 CopyOut(outGmOffset, 1, onceNum);
244 }243 }
@@ -261,7 +260,7 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::Process()
261 inQueX_.FreeTensor(inputUb);260 inQueX_.FreeTensor(inputUb);
262 LocalTensor<T> outUb = outQueY_.AllocTensor<T>();261 LocalTensor<T> outUb = outQueY_.AllocTensor<T>();
263 auto outUbAddr = (__ubuf__ T*)outUb.GetPhyAddr();262 auto outUbAddr = (__ubuf__ T*)outUb.GetPhyAddr();
264- UbDataCopy(swigluAddr, outUbAddr, onceBNum * alignDim1In);263+ UbStoreAlign(swigluAddr, outUbAddr, onceBNum * alignDim1In);
265 outQueY_.EnQue(outUb);264 outQueY_.EnQue(outUb);
266 int64_t outGmOffset = (bCoreOffset + bi * bUbFactor_) * dimH_ + hCoreOffset + hi * hUbFactor_;265 int64_t outGmOffset = (bCoreOffset + bi * bUbFactor_) * dimH_ + hCoreOffset + hi * hUbFactor_;
267 CopyOut(outGmOffset, onceBNum, onceHNum);266 CopyOut(outGmOffset, onceBNum, onceHNum);
@@ -295,9 +294,9 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::CopyIn(in
295}294}
296 295 
297template <typename T, bool isInterleaved, bool isGroup>296template <typename T, bool isInterleaved, bool isGroup>
298-__aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::UbDataCopy(__local_mem__ T* inAddr,297+__aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::UbStoreAlign(__ubuf__ T* inAddr,
299- __local_mem__ T* outAddr,298+ __ubuf__ T* outAddr,
300- int64_t onceNum)299+ int64_t onceNum)
301{300{
302 uint32_t size = onceNum;301 uint32_t size = onceNum;
303 uint32_t vfLen = vfLenT_;302 uint32_t vfLen = vfLenT_;
@@ -309,16 +308,16 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::UbDataCop
309 for (uint16_t i = 0; i < times; i++) {308 for (uint16_t i = 0; i < times; i++) {
310 mask = MicroAPI::UpdateMask<T>(size);309 mask = MicroAPI::UpdateMask<T>(size);
311 AscendC::MicroAPI::AddrReg srcIdxOffset = AscendC::MicroAPI::CreateAddrReg<T>(i, vfLen);310 AscendC::MicroAPI::AddrReg srcIdxOffset = AscendC::MicroAPI::CreateAddrReg<T>(i, vfLen);
312- AscendC::MicroAPI::DataCopy(xReg, inAddr, srcIdxOffset);311+ AscendC::MicroAPI::LoadAlign(xReg, inAddr, srcIdxOffset);
313- AscendC::MicroAPI::DataCopy(outAddr, xReg, srcIdxOffset, mask);312+ AscendC::MicroAPI::StoreAlign(outAddr, xReg, srcIdxOffset, mask);
314 }313 }
315 }314 }
316}315}
317 316 
318template <typename T, bool isInterleaved, bool isGroup>317template <typename T, bool isInterleaved, bool isGroup>
319__aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::ComputeVfSwiglu(318__aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::ComputeVfSwiglu(
320- __local_mem__ T* x1UbAddr, __local_mem__ T* x2UbAddr, __local_mem__ T* swigluUbAddr, int64_t dim0OnceSize,319+ __ubuf__ T* x1UbAddr, __ubuf__ T* x2UbAddr, __ubuf__ T* swigluUbAddr, int64_t dim0OnceSize, int64_t dim1OnceSize,
321- int64_t dim1OnceSize, int64_t alignDim1In)320+ int64_t alignDim1In)
322{321{
323 float clampLimit = limit_;322 float clampLimit = limit_;
324 float negClampLimit = -limit_;323 float negClampLimit = -limit_;
@@ -344,9 +343,9 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::ComputeVf
344 if (tail > 0) {343 if (tail > 0) {
345 tailTimes = 1;344 tailTimes = 1;
346 }345 }
347- __local_mem__ T* x1UbAddrT = x1UbAddr + dim1VfTimes * vfLen;346+ __ubuf__ T* x1UbAddrT = x1UbAddr + dim1VfTimes * vfLen;
348- __local_mem__ T* x2UbAddrT = x2UbAddr + dim1VfTimes * vfLen;347+ __ubuf__ T* x2UbAddrT = x2UbAddr + dim1VfTimes * vfLen;
349- __local_mem__ T* swigluUbAddrT = swigluUbAddr + dim1VfTimes * VF_LEN_FP32;348+ __ubuf__ T* swigluUbAddrT = swigluUbAddr + dim1VfTimes * VF_LEN_FP32;
350 __VEC_SCOPE__349 __VEC_SCOPE__
351 {350 {
352 AscendC::MicroAPI::RegTensor<T> vregX1;351 AscendC::MicroAPI::RegTensor<T> vregX1;
@@ -371,30 +370,30 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::ComputeVf
371 dim0vfLoopIdx, alignDim1In, dim1vfLoopIdx, vfLen);370 dim0vfLoopIdx, alignDim1In, dim1vfLoopIdx, vfLen);
372 if constexpr (isInterleaved) {371 if constexpr (isInterleaved) {
373 if constexpr (sizeof(T) == sizeof(half)) {372 if constexpr (sizeof(T) == sizeof(half)) {
374- AscendC::MicroAPI::DataCopy<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX1, x1UbAddr,373+ AscendC::MicroAPI::LoadAlign<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX1, x1UbAddr,
375- srcIdxOffset);374+ srcIdxOffset);
376- AscendC::MicroAPI::DataCopy<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX2, x2UbAddr,375+ AscendC::MicroAPI::LoadAlign<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX2, x2UbAddr,
377- srcIdxOffset);376+ srcIdxOffset);
378 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX1F, vregX1, mask);377 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX1F, vregX1, mask);
379 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX2F, vregX2, mask);378 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX2F, vregX2, mask);
380 } else {379 } else {
381 // float380 // float
382- AscendC::MicroAPI::DataCopy((MicroAPI::RegTensor<T>&)vregX1F, x1UbAddr, srcIdxOffset);381+ AscendC::MicroAPI::LoadAlign((MicroAPI::RegTensor<T>&)vregX1F, x1UbAddr, srcIdxOffset);
383- AscendC::MicroAPI::DataCopy((MicroAPI::RegTensor<T>&)vregX2F, x2UbAddr, srcIdxOffset);382+ AscendC::MicroAPI::LoadAlign((MicroAPI::RegTensor<T>&)vregX2F, x2UbAddr, srcIdxOffset);
384 }383 }
385 AscendC::MicroAPI::DeInterleave(vregX1DeF, vregX2DeF, vregX1F, vregX2F);384 AscendC::MicroAPI::DeInterleave(vregX1DeF, vregX2DeF, vregX1F, vregX2F);
386 } else {385 } else {
387 if constexpr (sizeof(T) == sizeof(half)) {386 if constexpr (sizeof(T) == sizeof(half)) {
388- AscendC::MicroAPI::DataCopy<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX1, x1UbAddr,387+ AscendC::MicroAPI::LoadAlign<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX1, x1UbAddr,
389- srcIdxOffset);388+ srcIdxOffset);
390- AscendC::MicroAPI::DataCopy<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX2, x2UbAddr,389+ AscendC::MicroAPI::LoadAlign<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX2, x2UbAddr,
391- srcIdxOffset);390+ srcIdxOffset);
392 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX1DeF, vregX1, mask);391 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX1DeF, vregX1, mask);
393 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX2DeF, vregX2, mask);392 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX2DeF, vregX2, mask);
394 } else {393 } else {
395 // float394 // float
396- AscendC::MicroAPI::DataCopy((MicroAPI::RegTensor<T>&)vregX1DeF, x1UbAddr, srcIdxOffset);395+ AscendC::MicroAPI::LoadAlign((MicroAPI::RegTensor<T>&)vregX1DeF, x1UbAddr, srcIdxOffset);
397- AscendC::MicroAPI::DataCopy((MicroAPI::RegTensor<T>&)vregX2DeF, x2UbAddr, srcIdxOffset);396+ AscendC::MicroAPI::LoadAlign((MicroAPI::RegTensor<T>&)vregX2DeF, x2UbAddr, srcIdxOffset);
398 }397 }
399 }398 }
400 AscendC::MicroAPI::Mins(minsReg, vregX1DeF, clampLimit, mask);399 AscendC::MicroAPI::Mins(minsReg, vregX1DeF, clampLimit, mask);
@@ -412,9 +411,9 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::ComputeVf
412 dim1vfLoopIdx, VF_LEN_FP32);411 dim1vfLoopIdx, VF_LEN_FP32);
413 if constexpr (sizeof(T) == sizeof(half)) {412 if constexpr (sizeof(T) == sizeof(half)) {
414 AscendC::MicroAPI::Cast<T, float, CAST_FP32_TO_FP16_BF16>(outTReg, outFReg, mask);413 AscendC::MicroAPI::Cast<T, float, CAST_FP32_TO_FP16_BF16>(outTReg, outFReg, mask);
415- DataCopy<T, AscendC::MicroAPI::StoreDist::DIST_PACK_B32>(swigluUbAddr, outTReg, outOffset, mask);414+ StoreAlign<T, AscendC::MicroAPI::StoreDist::DIST_PACK_B32>(swigluUbAddr, outTReg, outOffset, mask);
416 } else {415 } else {
417- DataCopy(swigluUbAddr, (MicroAPI::RegTensor<T>&)outFReg, outOffset, mask);416+ StoreAlign(swigluUbAddr, (MicroAPI::RegTensor<T>&)outFReg, outOffset, mask);
418 }417 }
419 }418 }
420 AscendC::MicroAPI::AddrReg srcIdxOffset1 = AscendC::MicroAPI::CreateAddrReg<T>(dim0vfLoopIdx, alignDim1In);419 AscendC::MicroAPI::AddrReg srcIdxOffset1 = AscendC::MicroAPI::CreateAddrReg<T>(dim0vfLoopIdx, alignDim1In);
@@ -422,30 +421,30 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::ComputeVf
422 for (uint16_t ti = 0; ti < tailTimes; ti++) {421 for (uint16_t ti = 0; ti < tailTimes; ti++) {
423 if constexpr (isInterleaved) {422 if constexpr (isInterleaved) {
424 if constexpr (sizeof(T) == sizeof(half)) {423 if constexpr (sizeof(T) == sizeof(half)) {
425- AscendC::MicroAPI::DataCopy<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX1, x1UbAddrT,424+ AscendC::MicroAPI::LoadAlign<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX1, x1UbAddrT,
426- srcIdxOffset1);425+ srcIdxOffset1);
427- AscendC::MicroAPI::DataCopy<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX2, x2UbAddrT,426+ AscendC::MicroAPI::LoadAlign<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX2, x2UbAddrT,
428- srcIdxOffset1);427+ srcIdxOffset1);
429 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX1F, vregX1, mask);428 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX1F, vregX1, mask);
430 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX2F, vregX2, mask);429 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX2F, vregX2, mask);
431 } else {430 } else {
432 // float431 // float
433- AscendC::MicroAPI::DataCopy((MicroAPI::RegTensor<T>&)vregX1F, x1UbAddrT, srcIdxOffset1);432+ AscendC::MicroAPI::LoadAlign((MicroAPI::RegTensor<T>&)vregX1F, x1UbAddrT, srcIdxOffset1);
434- AscendC::MicroAPI::DataCopy((MicroAPI::RegTensor<T>&)vregX2F, x2UbAddrT, srcIdxOffset1);433+ AscendC::MicroAPI::LoadAlign((MicroAPI::RegTensor<T>&)vregX2F, x2UbAddrT, srcIdxOffset1);
435 }434 }
436 AscendC::MicroAPI::DeInterleave(vregX1DeF, vregX2DeF, vregX1F, vregX2F);435 AscendC::MicroAPI::DeInterleave(vregX1DeF, vregX2DeF, vregX1F, vregX2F);
437 } else {436 } else {
438 if constexpr (sizeof(T) == sizeof(half)) {437 if constexpr (sizeof(T) == sizeof(half)) {
439- AscendC::MicroAPI::DataCopy<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX1, x1UbAddrT,438+ AscendC::MicroAPI::LoadAlign<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX1, x1UbAddrT,
440- srcIdxOffset1);439+ srcIdxOffset1);
441- AscendC::MicroAPI::DataCopy<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX2, x2UbAddrT,440+ AscendC::MicroAPI::LoadAlign<T, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(vregX2, x2UbAddrT,
442- srcIdxOffset1);441+ srcIdxOffset1);
443 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX1DeF, vregX1, mask);442 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX1DeF, vregX1, mask);
444 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX2DeF, vregX2, mask);443 AscendC::MicroAPI::Cast<float, T, CAST_BF16_FP16_TO_FP32>(vregX2DeF, vregX2, mask);
445 } else {444 } else {
446 // float445 // float
447- AscendC::MicroAPI::DataCopy((MicroAPI::RegTensor<T>&)vregX1DeF, x1UbAddrT, srcIdxOffset1);446+ AscendC::MicroAPI::LoadAlign((MicroAPI::RegTensor<T>&)vregX1DeF, x1UbAddrT, srcIdxOffset1);
448- AscendC::MicroAPI::DataCopy((MicroAPI::RegTensor<T>&)vregX2DeF, x2UbAddrT, srcIdxOffset1);447+ AscendC::MicroAPI::LoadAlign((MicroAPI::RegTensor<T>&)vregX2DeF, x2UbAddrT, srcIdxOffset1);
449 }448 }
450 }449 }
451 AscendC::MicroAPI::Mins(minsReg, vregX1DeF, clampLimit, maskT);450 AscendC::MicroAPI::Mins(minsReg, vregX1DeF, clampLimit, maskT);
@@ -461,9 +460,10 @@ __aicore__ inline void ClippedSwigluKernel<T, isInterleaved, isGroup>::ComputeVf
461 AscendC::MicroAPI::Mul(outFReg, sigmoidReg, vregX2DeF, maskT);460 AscendC::MicroAPI::Mul(outFReg, sigmoidReg, vregX2DeF, maskT);
462 if constexpr (sizeof(T) == sizeof(half)) {461 if constexpr (sizeof(T) == sizeof(half)) {
463 AscendC::MicroAPI::Cast<T, float, CAST_FP32_TO_FP16_BF16>(outTReg, outFReg, maskT);462 AscendC::MicroAPI::Cast<T, float, CAST_FP32_TO_FP16_BF16>(outTReg, outFReg, maskT);
464- DataCopy<T, AscendC::MicroAPI::StoreDist::DIST_PACK_B32>(swigluUbAddrT, outTReg, outOffset1, maskT);463+ StoreAlign<T, AscendC::MicroAPI::StoreDist::DIST_PACK_B32>(swigluUbAddrT, outTReg, outOffset1,
464+ maskT);
465 } else {465 } else {
466- DataCopy(swigluUbAddrT, (MicroAPI::RegTensor<T>&)outFReg, outOffset1, maskT);466+ StoreAlign(swigluUbAddrT, (MicroAPI::RegTensor<T>&)outFReg, outOffset1, maskT);
467 }467 }
468 }468 }
469 }469 }
@@ -2,10 +2,10 @@
2# -*- coding: utf-8 -*-2# -*- coding: utf-8 -*-
3# ----------------------------------------------------------------------------3# ----------------------------------------------------------------------------
4# Copyright (c) 2025 Huawei Technologies Co., Ltd.4# Copyright (c) 2025 Huawei Technologies Co., Ltd.
5-# This program is free software, you can redistribute it and/or modify it under the terms and conditions of 5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6# CANN Open Software License Agreement Version 2.0 (the "License").6# CANN Open Software License Agreement Version 2.0 (the "License").
7# Please refer to the License for details. You may not use this file except in compliance with the License.7# Please refer to the License for details. You may not use this file except in compliance with the License.
8-# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, 8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.9# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10# See LICENSE in the root of the software repository for the full text of the License.10# See LICENSE in the root of the software repository for the full text of the License.
11# ----------------------------------------------------------------------------11# ----------------------------------------------------------------------------
@@ -13,7 +13,8 @@
13import sys13import sys
14import numpy as np14import numpy as np
15import torch15import torch
16-import tensorflow as tf16+from ml_dtypes import bfloat16
17+ 
17 18 
18def do_clippedSwiglu(x, group_index, dim, alpha, limit, bias, interleaved):19def do_clippedSwiglu(x, group_index, dim, alpha, limit, bias, interleaved):
19 shape = x.shape20 shape = x.shape
@@ -25,7 +26,7 @@ def do_clippedSwiglu(x, group_index, dim, alpha, limit, bias, interleaved):
25 group = x.shape[0]26 group = x.shape[0]
26 if group_index is not None:27 if group_index is not None:
27 group = min(group_index.sum(), x.shape[0])28 group = min(group_index.sum(), x.shape[0])
28- 29+ 
29 x_tensor = x[:group]30 x_tensor = x[:group]
30 if interleaved:31 if interleaved:
31 x_glu = x_tensor[..., ::2]32 x_glu = x_tensor[..., ::2]
@@ -48,27 +49,27 @@ def do_clippedSwiglu(x, group_index, dim, alpha, limit, bias, interleaved):
48 49 
49 return res50 return res
50 51 
52+ 
51params_info = {53params_info = {
52 "test_case_bf16_shortH": {"x_shape": [3200, 5760], "x_dtype": torch.bfloat16},54 "test_case_bf16_shortH": {"x_shape": [3200, 5760], "x_dtype": torch.bfloat16},
53 "test_case_fp16_shortH": {"x_shape": [3200, 5760], "x_dtype": torch.float16},55 "test_case_fp16_shortH": {"x_shape": [3200, 5760], "x_dtype": torch.float16},
54 "test_case_fp32_shortH": {"x_shape": [3200, 5760], "x_dtype": torch.float32},56 "test_case_fp32_shortH": {"x_shape": [3200, 5760], "x_dtype": torch.float32},
55 "test_case_bf16_longH": {"x_shape": [3200, 23040], "x_dtype": torch.bfloat16},57 "test_case_bf16_longH": {"x_shape": [3200, 23040], "x_dtype": torch.bfloat16},
56 "test_case_fp16_longH": {"x_shape": [3200, 23040], "x_dtype": torch.float16},58 "test_case_fp16_longH": {"x_shape": [3200, 23040], "x_dtype": torch.float16},
57- "test_case_fp32_longH": {"x_shape": [3200, 23040], "x_dtype": torch.float32}59+ "test_case_fp32_longH": {"x_shape": [3200, 23040], "x_dtype": torch.float32},
58- }60+}
59 61 
60 62 
61def gen_golden_data_simple(params_list):63def gen_golden_data_simple(params_list):
62 # np.random.seed(1)64 # np.random.seed(1)
63 x_shape = params_list["x_shape"]65 x_shape = params_list["x_shape"]
64- x_dtype = params_list["x_dtype"]
65 data_x = torch.randn(x_shape, dtype=torch.float32)66 data_x = torch.randn(x_shape, dtype=torch.float32)
66 group_index = torch.randint(low=1, high=11, size=(10,))67 group_index = torch.randint(low=1, high=11, size=(10,))
67- out_list = do_clippedSwiglu(data_x, group_index, -1, 1.702, 7.0, 1.0, True)68+ do_clippedSwiglu(data_x, group_index, -1, 1.702, 7.0, 1.0, True)
68- 69+ 
69 data_x = data_x.cpu().numpy()70 data_x = data_x.cpu().numpy()
70 if params_list["x_dtype"] == torch.bfloat16:71 if params_list["x_dtype"] == torch.bfloat16:
71- data_x = data_x.astype(tf.bfloat16.as_numpy_dtype)72+ data_x = data_x.astype(bfloat16)
72 elif params_list["x_dtype"] == torch.float16:73 elif params_list["x_dtype"] == torch.float16:
73 data_x = data_x.astype(np.float16)74 data_x = data_x.astype(np.float16)
74 elif params_list["x_dtype"] == torch.float32:75 elif params_list["x_dtype"] == torch.float32:
@@ -77,9 +78,6 @@ def gen_golden_data_simple(params_list):
77 group_index.cpu().numpy().tofile("./input_group_index.bin")78 group_index.cpu().numpy().tofile("./input_group_index.bin")
78 79 
79 80 
80- 
81- 
82- 
83if __name__ == "__main__":81if __name__ == "__main__":
84 params_list = params_info[sys.argv[1]]82 params_list = params_info[sys.argv[1]]
85- gen_golden_data_simple(params_list)83+ gen_golden_data_simple(params_list)
@@ -0,0 +1,20 @@
1+# ----------------------------------------------------------------------------
2+# Copyright (c) 2025 Huawei Technologies Co., Ltd.
3+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+# CANN Open Software License Agreement Version 2.0 (the "License").
5+# Please refer to the License for details. You may not use this file except in compliance with the License.
6+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+# See LICENSE in the root of the software repository for the full text of the License.
9+# ----------------------------------------------------------------------------
10+ 
11+message(STATUS "=== Debug: CURRENT_DIRS =${CURRENT_DIRS} ")
12+file(GLOB CURRENT_DIRS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
13+if(NOT ENABLE_TEST AND NOT BENCHMARK)
14+ list(REMOVE_ITEM CURRENT_DIRS tests)
15+endif()
16+foreach(SUB_DIR ${CURRENT_DIRS})
17+ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
18+ add_subdirectory(${SUB_DIR})
19+ endif()
20+endforeach()
@@ -0,0 +1,67 @@
1+# ClippedSwigluGrad
2+ 
3+## 产品支持情况
4+ 
5+|产品 | 是否支持 |
6+|:-------------------------|:----------:|
7+| <term>Ascend 950PR/Ascend 950DT</term> | √ |
8+| <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term> | √ |
9+| <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term> | √ |
10+| <term>Atlas 200I/500 A2 推理产品</term> | × |
11+| <term>Atlas 推理系列产品</term> | × |
12+| <term>Atlas 训练系列产品</term> | × |
13+ 
14+## 功能说明
15+ 
16+- 接口功能:aclnnClippedSwiglu的反向算子,计算输入x的梯度`gradXOut`。
17+ 
18+- 计算公式:
19+ 
20+ 对给定的正向输入张量x(维度为[a,b,c,d,e,f,g…])和输出梯度gradY,ClippedSwigluGrad对其进行以下计算:
21+ 
22+ 1. 将x基于输入参数dim进行合轴,合轴后维度为[pre, dim2H];gradY合轴后维度为[pre, dim2H//2]。
23+ 
24+ 2. 根据输入参数group_index,对x的pre轴进行过滤处理,仅前group行参与计算,其余行梯度为0。
25+ 
26+ 3. 根据输入参数interleaved,对x进行切分,得到a和b两部分(与正向一致)。
27+ 
28+ 4. 重算正向中间量:
29+ - A = clamp(a, max=limit)
30+ - B = clamp(b, -limit, limit)
31+ - s = sigmoid(alpha * A)
32+ 
33+ 5. 计算反向梯度:
34+ - maskX0 = (a <= limit)
35+ - maskX1 = (-limit <= b <= limit)
36+ - da = gradY * (B + bias) * s * (1 + alpha * A * (1 - s)) * maskX0
37+ - db = gradY * A * s * maskX1
38+ 
39+ 6. 将da、db散回dx:
40+ - interleaved为true时:dx[..., ::2] = da,dx[..., 1::2] = db
41+ - interleaved为false时:dx[..., :h] = da,dx[..., h:] = db
42+ 
43+ 7. 重塑输出张量grad_x的维度与x一致。
44+ 
45+## 参数说明
46+ 
47+|参数名|输入/输出/属性|描述|数据类型|数据格式|
48+|---|---|---|---|---|
49+|grad_y|输入|正向输出y的梯度。维度在dim轴上为x的一半,其他维度与x一致。|FLOAT、FLOAT16、BFLOAT16|ND|
50+|x|输入|正向输入x。|FLOAT、FLOAT16、BFLOAT16|ND|
51+|group_index|可选输入|分组索引。1维,元素个数不超过8192。|INT64|-|
52+|dim|可选属性|合轴以及切分的维度序号。取值范围[-x.dim(), x.dim()-1]。默认-1。|INT64|-|
53+|alpha|可选属性|SwiGLU激活系数。默认1.702。|FLOAT|-|
54+|limit|可选属性|截断门限值。必须大于0。默认7.0。|FLOAT|-|
55+|bias|可选属性|线性计算偏差。默认1.0。|FLOAT|-|
56+|interleaved|可选属性|切分方式。true=奇偶切分,false=前后切分。默认true。|BOOL|-|
57+|grad_x|输出|x的梯度,与x同shape。|FLOAT、FLOAT16、BFLOAT16|ND|
58+ 
59+## 约束说明
60+ 
61+无。
62+ 
63+## 调用说明
64+ 
65+| 调用方式 | 调用样例 | 说明 |
66+|---|---|---|
67+| aclnn调用 | [test_aclnn_clipped_swiglu_grad](./examples/test_aclnn_clipped_swiglu_grad.cpp) | 通过[aclnnClippedSwigluGrad](./docs/aclnnClippedSwigluGrad.md)接口方式调用ClippedSwigluGrad算子。 |
@@ -0,0 +1,539 @@
1+# aclnnClippedSwigluGrad
2+ 
3+[📄 查看源码](https://gitcode.com/cann/ops-nn/tree/master/activation/clipped_swiglu_grad)
4+ 
5+## 产品支持情况
6+ 
7+<!-- npu="950" id1 -->
8+- <term>Ascend 950PR/Ascend 950DT</term>:支持
9+<!-- end id1 -->
10+<!-- npu="A3" id2 -->
11+- <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term>:支持
12+<!-- end id2 -->
13+<!-- npu="910b" id3 -->
14+- <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term>:支持
15+<!-- end id3 -->
16+<!-- npu="310b" id4 -->
17+- <term>Atlas 200I/500 A2 推理产品</term>:不支持
18+<!-- end id4 -->
19+<!-- npu="310p" id5 -->
20+- <term>Atlas 推理系列产品</term>:不支持
21+<!-- end id5 -->
22+<!-- npu="910" id6 -->
23+- <term>Atlas 训练系列产品</term>:不支持
24+<!-- end id6 -->
25+ 
26+## 功能说明
27+ 
28+- 接口功能:aclnnClippedSwiglu的反向算子,计算输入x的梯度`gradXOut`。
29+ 
30+- 计算流程:
31+ - 步骤一:合轴与切分(将x合轴并切分为x0和x1)
32+ - 步骤二:GroupIndex处理(可选)→ 计算有效行数
33+ - 步骤三:重算正向中间量
34+ - 步骤四:Clamp反向传播掩码计算
35+ - 步骤五:SwiGLU反向梯度计算
36+ - 步骤六:梯度拼接输出
37+ - 步骤七:GroupIndex置零处理(可选)
38+ 
39+- 合轴与切分公式:
40+ 
41+ 将x基于输入参数`dim`进行合轴,合轴后维度为[pre, cut, after]。其中cut轴为需要切分的轴,pre和after可以等于1。由于after轴的元素为连续存放且计算操作为逐元素的,将cut轴与after轴合并,得到x的维度为[pre, cut]。`gradY`合轴后维度为[pre, cut/2]。
42+ 
43+ 根据`interleaved`参数对x进行切分:
44+ 
45+ 当`interleaved`为true时,表示奇偶切分:
46+ 
47+ $$
48+ \mathbf{x}_0[p, h] = \mathbf{x}[p, 2h], \quad h \in [0, \text{cut}/2)
49+ $$
50+ 
51+ $$
52+ \mathbf{x}_1[p, h] = \mathbf{x}[p, 2h+1], \quad h \in [0, \text{cut}/2)
53+ $$
54+ 
55+ 当`interleaved`为false时,表示前后切分:
56+ 
57+ $$
58+ H = \text{cut} / 2
59+ $$
60+ 
61+ $$
62+ \mathbf{x}_0[p, h] = \mathbf{x}[p, h], \quad h \in [0, H)
63+ $$
64+ 
65+ $$
66+ \mathbf{x}_1[p, h] = \mathbf{x}[p, h + H], \quad h \in [0, H)
67+ $$
68+ 
69+ 其中$p$为pre维索引,$h$为切分后半宽索引。
70+ 
71+- GroupIndex处理公式(当提供`groupIndex`时):
72+ 
73+ $$
74+ \text{trunc} = \sum_{g=0}^{G-1} \text{groupIndex}[g]
75+ $$
76+ 
77+ $$
78+ \text{validRows} = \min(\text{trunc}, \text{pre})
79+ $$
80+ 
81+ 其中:$G$为groupIndex的元素个数,后续所有步骤仅处理前$\text{validRows}$行数据。当不提供`groupIndex`时,$\text{validRows} = \text{pre}$。
82+ 
83+- 重算正向中间量公式:
84+ 
85+ $$
86+ \mathbf{x}_0' = \min(\mathbf{x}_0, l)
87+ $$
88+ 
89+ $$
90+ \mathbf{x}_1' = \min(\max(\mathbf{x}_1, -l), l)
91+ $$
92+ 
93+ $$
94+ s = \sigma(\alpha \times \mathbf{x}_0') = \frac{1}{1 + e^{-\alpha \times \mathbf{x}_0'}}
95+ $$
96+ 
97+ 其中:$\mathbf{x}_0'$为对$\mathbf{x}_0$仅做上界截断,$\mathbf{x}_1'$为对$\mathbf{x}_1$做双向截断,$s$为sigmoid激活值,$\alpha$为`alpha`参数,$l$为`limit`参数。
98+ 
99+- Clamp反向传播掩码公式:
100+ 
@@ -0,0 +1,131 @@
1+# clipped_swiglu_grad
2+ 
3+## 产品支持情况
4+ 
5+<!-- npu="950" id1 -->
6+- <term>Ascend 950PR/Ascend 950DT</term>:支持
7+<!-- end id1 -->
8+<!-- npu="A3" id2 -->
9+- <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term>:支持
10+<!-- end id2 -->
11+<!-- npu="910b" id3 -->
12+- <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term>:支持
13+<!-- end id3 -->
14+<!-- npu="310b" id4 -->
15+- <term>Atlas 200I/500 A2 推理产品</term>:不支持
16+<!-- end id4 -->
17+<!-- npu="310p" id5 -->
18+- <term>Atlas推理系列产品</term>:不支持
19+<!-- end id5 -->
20+<!-- npu="910" id6 -->
21+- <term>Atlas训练系列产品</term>:不支持
22+<!-- end id6 -->
23+ 
24+## 功能说明
25+ 
26+- 接口功能:
27+ 
28+ `ClippedSwigluGrad`是`ClippedSwiglu`算子的反向梯度计算算子。根据正向输出`y`的梯度`grad_y`、前向输入`x`以及可选的分组索引`group_index`,反向计算输入`x`的梯度`grad_x`。底层封装`aclnnClippedSwigluGrad`。
29+ 
30+- 计算流程:
31+ 
32+ 1. 将`x`基于参数`dim`进行合轴,合轴后维度为`[pre, dim2H]`;`grad_y`合轴后维度为`[pre, dim2H // 2]`。
33+ 2. 根据`group_index`,对`x`的`pre`轴进行过滤处理,仅前`group`行参与计算,其余行梯度为 0。
34+ 3. 根据`interleaved`,对`x`进行切分,得到`x0`和`x1`两部分(与正向一致)。
35+ 4. 重算正向中间量:
36+ - $\mathbf{x}_0' = \min(\mathbf{x}_0, \text{limit})$
37+ - $\mathbf{x}_1' = \min(\max(\mathbf{x}_1, -\text{limit}), \text{limit})$
38+ - $s = \sigma(\alpha \times \mathbf{x}_0') = \frac{1}{1 + e^{-\alpha \times \mathbf{x}_0'}}$
39+ 5. 计算反向梯度:
40+ - $\text{mask}_{x_0} = (\mathbf{x}_0 \le \text{limit})$
41+ - $\text{mask}_{x_1} = (-\text{limit} \le \mathbf{x}_1 \le \text{limit})$
42+ - $\mathbf{grad}_{x_0} = \text{grad\_y} \times (\mathbf{x}_1' + \text{bias}) \times s \times (1 + \alpha \times \mathbf{x}_0' \times (1 - s)) \times \text{mask}_{x_0}$
43+ - $\mathbf{grad}_{x_1} = \text{grad\_y} \times \mathbf{x}_0' \times s \times \text{mask}_{x_1}$
44+ 6. 将`grad_x0`、`grad_x1`散回`dx`(与切分方式对应),重塑输出维度与`x`一致。
45+ 
46+## 函数原型
47+ 
48+```python
49+cann_ops_nn.clipped_swiglu_grad(
50+ grad_y,
51+ x,
52+ *,
53+ group_index=None,
54+ dim=-1,
55+ alpha=1.702,
56+ limit=7.0,
57+ bias=1.0,
58+ interleaved=True,
59+ clamp_mode=0,
60+) -> Tensor
61+```
62+ 
63+## 参数说明
64+ 
65+| 参数名 | 参数类型 | 可选/必选 | 描述 | 数据类型 | 维度(shape) |
66+| --- | --- | --- | --- | --- | --- |
67+| `grad_y` | Tensor | 必选 | 正向输出`y`的梯度,`dim`维为`x`的一半。 | `torch.float16`、`torch.bfloat16`、`torch.float32` | 与`x`相同,但`dim`维为`x`对应维度的一半 |
68+| `x` | Tensor | 必选 | 正向输入,用于重算正向中间量。 | `torch.float16`、`torch.bfloat16`、`torch.float32` | 1-8维,`dim`维大小为偶数 |
69+| `group_index` | Tensor | 可选 | 分组索引,第`i`个元素表示第`i`组处理的batch数量。 | `torch.int64` | 1维,元素个数不超过 8192 |
70+| `dim` | int | 可选 | 合轴及切分的维度序号,取值范围`[-x.dim(), x.dim()-1]`,默认`-1`。 | - | - |
71+| `alpha` | float | 可选 | SwiGLU激活系数,默认`1.702`。 | - | - |
72+| `limit` | float | 可选 | 截断门限值,必须大于 0,默认`7.0`。 | - | - |
73+| `bias` | float | 可选 | 线性计算偏差,默认`1.0`。 | - | - |
74+| `interleaved` | bool | 可选 | 切分方式。`true`=奇偶切分,`false`=前后切分,默认`true`。 | - | - |
75+| `clamp_mode` | int | 可选 | clamp位置控制,默认`0`。当前仅支持`0`(clamp在silu之前),后续版本将支持`1`(clamp移至silu之后)。 | - | - |
76+ 
77+## 返回值说明
78+ 
79+| 参数名 | 参数类型 | 描述 | 数据类型 | 维度(shape) |
80+| --- | --- | --- | --- | --- |
81+| `grad_x` | Tensor | 输入`x`的梯度。 | 与`x`相同 | 与`x`相同 |
82+ 
83+## 约束说明
84+ 
85+- 该接口支持单算子模式和TorchAir图模式调用。
86+- `grad_y`、`x`均需为NPU Tensor;`group_index`可传`None`。
87+- `grad_y`与`x`的数据类型必须一致。
88+- `x`的rank必须大于 0,`dim`维大小必须为偶数。
89+- `grad_y`的`dim`维大小必须等于`x`对应维度的一半,其余维度与`x`一致。
90+- `limit`必须大于 0。
91+- `clamp_mode`当前仅支持取值`0`。
92+- 不支持空Tensor和非连续Tensor。
93+ 
94+## 确定性计算
95+ 
96+默认支持确定性计算。
97+ 
98+## 调用说明
99+ 
100+- 单算子模式调用:
101+ 
102+ ```python
103+ import torch
104+ import torch_npu
105+ import cann_ops_nn.ops
106+ 
107+ x = torch.randn(8, 512, dtype=torch.float16).npu()
108+ grad_y = torch.randn(8, 256, dtype=torch.float16).npu()
109+ 
110+ grad_x = torch.ops.cann_ops_nn.clipped_swiglu_grad(
111+ grad_y, x, dim=-1, alpha=1.702, limit=7.0, bias=1.0, interleaved=True)
112+ print(grad_x.shape, grad_x.dtype)
113+ ```
114+ 
115+- 图模式(torchair)调用:
116+ 
117+ ```python
118+ import torch
119+ import torch_npu
120+ import torchair
121+ import cann_ops_nn.ops
122+ 
123+ class Model(torch.nn.Module):
124+ def forward(self, grad_y, x):
125+ return torch.ops.cann_ops_nn.clipped_swiglu_grad(grad_y, x)
126+ 
127+ model = torch.compile(Model().npu(), backend=torchair.get_npu_backend(), dynamic=False)
128+ grad_y = torch.randn(8, 256, dtype=torch.float16).npu()
129+ x = torch.randn(8, 512, dtype=torch.float16).npu()
130+ grad_x = model(grad_y, x)
131+ ```
@@ -0,0 +1,143 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+ 
11+#include <iostream>
12+#include <vector>
13+#include "acl/acl.h"
14+#include "aclnnop/aclnn_clipped_swiglu_grad.h"
15+ 
16+#define CHECK_RET(cond, return_expr) \
17+ do { \
18+ if (!(cond)) { \
19+ return_expr; \
20+ } \
21+ } while (0)
22+ 
23+#define LOG_PRINT(message, ...) \
24+ do { \
25+ printf(message, ##__VA_ARGS__); \
26+ } while (0)
27+ 
28+int64_t GetShapeSize(const std::vector<int64_t>& shape)
29+{
30+ int64_t shapeSize = 1;
31+ for (auto i : shape) {
32+ shapeSize *= i;
33+ }
34+ return shapeSize;
35+}
36+ 
37+int Init(int32_t deviceId, aclrtStream* stream)
38+{
39+ auto ret = aclInit(nullptr);
40+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
41+ ret = aclrtSetDevice(deviceId);
42+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
43+ ret = aclrtCreateStream(stream);
44+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
45+ return 0;
46+}
47+ 
48+template <typename T>
49+int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
50+ aclDataType dataType, aclTensor** tensor)
51+{
52+ auto size = GetShapeSize(shape) * sizeof(T);
53+ auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
54+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
55+ ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
56+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);
57+ 
58+ std::vector<int64_t> strides(shape.size(), 1);
59+ for (int64_t i = shape.size() - 2; i >= 0; i--) {
60+ strides[i] = shape[i + 1] * strides[i + 1];
61+ }
62+ 
63+ *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
64+ shape.data(), shape.size(), *deviceAddr);
65+ return 0;
66+}
67+ 
68+int main()
69+{
70+ int32_t deviceId = 0;
71+ aclrtStream stream;
72+ auto ret = Init(deviceId, &stream);
73+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
74+ 
75+ // x shape [2, 32], grad_y shape [2, 16], grad_x shape [2, 32]
76+ std::vector<int64_t> xShape = {2, 32};
77+ std::vector<int64_t> gradYShape = {2, 16};
78+ std::vector<int64_t> gradXOutShape = {2, 32};
79+ void* xDeviceAddr = nullptr;
80+ void* gradYDeviceAddr = nullptr;
81+ void* gradXOutDeviceAddr = nullptr;
82+ aclTensor* x = nullptr;
83+ aclTensor* gradY = nullptr;
84+ aclTensor* gradXOut = nullptr;
85+ 
86+ std::vector<float> xHostData(64, 1.0f);
87+ std::vector<float> gradYHostData(32, 1.0f);
88+ std::vector<float> gradXOutHostData(64, 0.0f);
89+ 
90+ int dim = -1;
91+ float alpha = 1.702;
92+ float limit = 7.0;
93+ float bias = 1.0;
94+ bool interleaved = true;
95+ 
96+ ret = CreateAclTensor(xHostData, xShape, &xDeviceAddr, aclDataType::ACL_FLOAT, &x);
97+ CHECK_RET(ret == ACL_SUCCESS, return ret);
98+ ret = CreateAclTensor(gradYHostData, gradYShape, &gradYDeviceAddr, aclDataType::ACL_FLOAT, &gradY);
99+ CHECK_RET(ret == ACL_SUCCESS, return ret);
100+ ret = CreateAclTensor(gradXOutHostData, gradXOutShape, &gradXOutDeviceAddr, aclDataType::ACL_FLOAT, &gradXOut);
101+ CHECK_RET(ret == ACL_SUCCESS, return ret);
102+ 
103+ uint64_t workspaceSize = 0;
104+ aclOpExecutor* executor;
105+ ret = aclnnClippedSwigluGradGetWorkspaceSize(gradY, x, nullptr, dim, alpha, limit, bias, interleaved, gradXOut,
106+ &workspaceSize, &executor);
107+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnClippedSwigluGradGetWorkspaceSize failed. ERROR: %d\n", ret);
108+ return ret);
109+ 
110+ void* workspaceAddr = nullptr;
111+ if (workspaceSize > 0) {
112+ ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
113+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
114+ }
115+ ret = aclnnClippedSwigluGrad(workspaceAddr, workspaceSize, executor, stream);
116+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnClippedSwigluGrad failed. ERROR: %d\n", ret); return ret);
117+ 
118+ ret = aclrtSynchronizeStream(stream);
119+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
120+ 
121+ auto size = GetShapeSize(gradXOutShape);
122+ std::vector<float> resultData(size, 0);
123+ ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), gradXOutDeviceAddr,
124+ size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
125+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
126+ for (int64_t i = 0; i < size; i++) {
127+ LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
128+ }
129+ 
130+ aclDestroyTensor(x);
131+ aclDestroyTensor(gradY);
132+ aclDestroyTensor(gradXOut);
133+ aclrtFree(xDeviceAddr);
134+ aclrtFree(gradYDeviceAddr);
135+ aclrtFree(gradXOutDeviceAddr);
136+ if (workspaceSize > 0) {
137+ aclrtFree(workspaceAddr);
138+ }
139+ aclrtDestroyStream(stream);
140+ aclrtResetDevice(deviceId);
141+ aclFinalize();
142+ return 0;
143+}
@@ -0,0 +1,11 @@
1+# ----------------------------------------------------------------------------
2+# Copyright (c) 2025 Huawei Technologies Co., Ltd.
3+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+# CANN Open Software License Agreement Version 2.0 (the "License").
5+# Please refer to the License for details. You may not use this file except in compliance with the License.
6+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+# See LICENSE in the root of the software repository for the full text of the License.
9+# ----------------------------------------------------------------------------
10+message(STATUS "=== Debug: start ops.fusedop.clipped_swiglu_grad.op_graph.CMakeLists.txt ")
11+add_graph_plugin_sources()
@@ -0,0 +1,64 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+ 
11+/*!
12+ * \file clipped_swiglu_grad_proto.h
13+ * \brief
14+ */
15+ 
16+#ifndef OPS_BUILT_IN_OP_PROTO_INC_CLIPPED_SWIGLU_GRAD_PROTO_H_
17+#define OPS_BUILT_IN_OP_PROTO_INC_CLIPPED_SWIGLU_GRAD_PROTO_H_
18+ 
19+#include "graph/operator_reg.h"
20+ 
21+namespace ge {
22+/**
23+* @brief Compute the ClippedSwigluGrad,
24+* where the activations function in GLU is SwishGrad.
25+ 
26+* @par Inputs:
27+* Three inputs, including:
28+* @li grad_y: A Tensor, which is the output gradient of forward operator and which
29+* has the same shape as "x" except for the dimension specified by the "dim" parameter.
30+* The dimension size specified by "dim" is half of the corresponding dimension of x.
31+* Must be one of the following types: bfloat16, float16, float32.
32+* @li x: A Tensor. Must be one of the following types: bfloat16, float16, float32.
33+* @li group_index: An optional tensor. Shape is (N,). Type is int64.
34+ 
35+* @par Outputs:
36+* one Output, including:
37+* grad_x: A Tensor, which is the gradient of x and has the same shape as "x".
38+* Must be one of the following types: bfloat16, float16, float32.
39+ 
40+* @par Attributes:
41+* Five attributes, including:
42+* @li dim: An optional int. The dimension to be split, value in [-xDim, xDim-1], default is -1.
43+* @li alpha: An optional float. The activation coefficient for the GLU activation function, default is 1.702.
44+* @li limit: An optional float. The threshold limit for SWIGLU input, default is 7.0.
45+* @li bias: An optional float. The bias applied during SWIGLU linear computation, default is 1.0.
46+ * @li interleaved: An optional bool. The way of splitting x: true for interleaved splitting,
47+ * false for front-back splitting, default is true.
48+* @par Attention Constraints:
49+* The dim dimension of x must be divisible by 2, and the dim dimension of grad_y must be equal
50+* to the dim dimension of x divided by 2.
51+*/
52+REG_OP(ClippedSwigluGrad)
53+ .INPUT(grad_y, TensorType({DT_BF16, DT_FLOAT16, DT_FLOAT}))
54+ .INPUT(x, TensorType({DT_BF16, DT_FLOAT16, DT_FLOAT}))
55+ .OPTIONAL_INPUT(group_index, TensorType({DT_INT64}))
56+ .OUTPUT(grad_x, TensorType({DT_BF16, DT_FLOAT16, DT_FLOAT}))
57+ .ATTR(dim, Int, -1)
58+ .ATTR(alpha, Float, 1.702)
59+ .ATTR(limit, Float, 7.0)
60+ .ATTR(bias, Float, 1.0)
61+ .ATTR(interleaved, Bool, true)
62+ .OP_END_FACTORY_REG(ClippedSwigluGrad)
63+} // namespace ge
64+#endif // OPS_BUILT_IN_OP_PROTO_INC_CLIPPED_SWIGLU_GRAD_PROTO_H_
Rrnn/dynamic_rnn/op_host/CMakeLists.txt→activation/clipped_swiglu_grad/op_host/CMakeLists.txt+6-3
@@ -1,9 +1,12 @@
1+# ----------------------------------------------------------------------------
1# Copyright (c) 2025 Huawei Technologies Co., Ltd.2# Copyright (c) 2025 Huawei Technologies Co., Ltd.
2-# This program is free software, you can redistribute it and/or modify it under the terms and conditions of 3+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
3# CANN Open Software License Agreement Version 2.0 (the "License").4# CANN Open Software License Agreement Version 2.0 (the "License").
4# Please refer to the License for details. You may not use this file except in compliance with the License.5# Please refer to the License for details. You may not use this file except in compliance with the License.
5-# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, 6+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
6# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.7# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
7# See LICENSE in the root of the software repository for the full text of the License.8# See LICENSE in the root of the software repository for the full text of the License.
9+# ----------------------------------------------------------------------------
8 10 
9-add_modules_sources(HOSTNAME ${OPHOST_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR} OPTYPE dynamic_rnn ACLNNTYPE aclnn_exclude)11+message(STATUS "=== Debug: CURRENT_DIRS =${CURRENT_DIRS} ")
12+add_modules_sources(HOSTNAME ${OPHOST_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR} OPTYPE clipped_swiglu_grad ACLNNTYPE aclnn DEPENDENCIES clipped_swiglu)
@@ -0,0 +1,74 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+ 
11+/*!
12+ * \file clipped_swiglu_grad_def.cpp
13+ * \brief
14+ */
15+#include <register/op_def_registry.h>
16+ 
17+namespace ops {
18+constexpr float DEFAULT_ALPHA = 1.702;
19+constexpr float DEFAULT_LIMIT = 7.0;
20+ 
21+static const std::vector<ge::DataType> xDtype = {ge::DT_FLOAT16, ge::DT_FLOAT, ge::DT_BF16};
22+static const std::vector<ge::DataType> groupIndexDtype = {ge::DT_INT64, ge::DT_INT64, ge::DT_INT64};
23+static const std::vector<ge::Format> xFormat = {ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND};
24+ 
25+class ClippedSwigluGrad : public OpDef {
26+public:
27+ explicit ClippedSwigluGrad(const char* name) : OpDef(name)
28+ {
29+ this->Input("grad_y").ParamType(REQUIRED).DataType(xDtype).Format(xFormat).UnknownShapeFormat(xFormat);
30+ this->Input("x").ParamType(REQUIRED).DataType(xDtype).Format(xFormat).UnknownShapeFormat(xFormat);
31+ this->Input("group_index")
32+ .ParamType(OPTIONAL)
33+ .DataType(groupIndexDtype)
34+ .Format(xFormat)
35+ .UnknownShapeFormat(xFormat);
36+ this->Output("grad_x").ParamType(REQUIRED).DataType(xDtype).Format(xFormat).UnknownShapeFormat(xFormat);
37+ this->Attr("dim").AttrType(OPTIONAL).Int(-1);
38+ this->Attr("alpha").AttrType(OPTIONAL).Float(DEFAULT_ALPHA);
39+ this->Attr("limit").AttrType(OPTIONAL).Float(DEFAULT_LIMIT);
40+ this->Attr("bias").AttrType(OPTIONAL).Float(1.0);
41+ this->Attr("interleaved").AttrType(OPTIONAL).Bool(true);
42+ 
43+ this->AICore().AddConfig("ascend910b");
44+ this->AICore().AddConfig("ascend910_93");
45+ 
46+ OpAICoreConfig regbaseConfig;
47+ regbaseConfig.Input("grad_y")
48+ .ParamType(REQUIRED)
49+ .DataType(xDtype)
50+ .Format(xFormat)
51+ .UnknownShapeFormat(xFormat)
52+ .AutoContiguous();
53+ regbaseConfig.Input("x")
54+ .ParamType(REQUIRED)
55+ .DataType(xDtype)
56+ .Format(xFormat)
57+ .UnknownShapeFormat(xFormat)
58+ .AutoContiguous();
59+ regbaseConfig.Input("group_index")
60+ .ParamType(OPTIONAL)
61+ .DataType(groupIndexDtype)
62+ .Format(xFormat)
63+ .UnknownShapeFormat(xFormat)
64+ .AutoContiguous();
65+ regbaseConfig.Output("grad_x").ParamType(REQUIRED).DataType(xDtype).Format(xFormat).UnknownShapeFormat(xFormat);
66+ regbaseConfig.DynamicCompileStaticFlag(true)
67+ .DynamicRankSupportFlag(true)
68+ .DynamicShapeSupportFlag(true)
69+ .ExtendCfgInfo("opFile.value", "clipped_swiglu_grad_apt");
70+ this->AICore().AddConfig("ascend950", regbaseConfig);
71+ }
72+};
73+OP_ADD(ClippedSwigluGrad);
74+} // namespace ops
@@ -0,0 +1,79 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+ 
11+/*!
12+ * \file clipped_swiglu_grad_infershape.cpp
13+ * \brief
14+ */
15+ 
16+#include "log/log.h"
17+#include "register/op_impl_registry.h"
18+#include "error_util.h"
19+#include "util/shape_util.h"
20+using namespace ge;
21+ 
22+namespace {
23+constexpr size_t GRAD_IN_X = 1; // x is the second input (index 1)
24+constexpr size_t GRAD_OUT_GRAD_X = 0; // grad_x is the first output (index 0)
25+constexpr size_t GRAD_ATTR_DIM = 0;
26+const size_t SPLIT_NUM = 2;
27+} // namespace
28+ 
29+namespace ops {
30+static ge::graphStatus InferShapeForClippedSwigluGrad(gert::InferShapeContext* context)
31+{
32+ OP_LOGD(context->GetNodeName(), "Begin to do InferShapeForClippedSwigluGrad");
33+ auto xShape = context->GetInputShape(GRAD_IN_X);
34+ OPS_CHECK_NULL_WITH_CONTEXT(context, xShape);
35+ auto gradXOutShape = context->GetOutputShape(GRAD_OUT_GRAD_X);
36+ OPS_CHECK_NULL_WITH_CONTEXT(context, gradXOutShape);
37+ auto attrs = context->GetAttrs();
38+ OPS_CHECK_NULL_WITH_CONTEXT(context, attrs);
39+ 
40+ auto splitDimPtr = attrs->GetAttrPointer<int64_t>(GRAD_ATTR_DIM);
41+ OPS_CHECK_NULL_WITH_CONTEXT(context, splitDimPtr);
42+ if (Ops::Base::IsUnknownRank(*xShape)) {
43+ Ops::Base::SetUnknownRank(*gradXOutShape);
44+ return ge::GRAPH_SUCCESS;
45+ }
46+ auto splitDim = *splitDimPtr;
47+ if (splitDim < 0) {
48+ splitDim += xShape->GetDimNum();
49+ }
50+ if (splitDim < 0 || splitDim >= static_cast<int64_t>(xShape->GetDimNum())) {
51+ OP_LOGE("ClippedSwigluGrad", "The value of attr [dim] must be in the range [-%zu, %zu], but got [%ld].",
52+ xShape->GetDimNum(), xShape->GetDimNum() - 1, splitDim);
53+ return GRAPH_FAILED;
54+ }
55+ int64_t splitDimSize = xShape->GetDim(splitDim);
56+ if (splitDimSize >= 0 && splitDimSize % static_cast<int64_t>(SPLIT_NUM) != 0) {
57+ OP_LOGE("ClippedSwigluGrad", "The size of x[dim] must be divisible by %zu, but got [%ld].", SPLIT_NUM,
58+ splitDimSize);
59+ return GRAPH_FAILED;
60+ }
61+ OP_LOGD(context->GetNodeName(), "Begin to generate gradXOutShape");
62+ // grad_x has the same shape as x
63+ *gradXOutShape = *xShape;
64+ return ge::GRAPH_SUCCESS;
65+}
66+ 
67+static ge::graphStatus InferDataTypeForClippedSwigluGrad(gert::InferDataTypeContext* context)
68+{
69+ OP_LOGD(context->GetNodeName(), "Begin to do InferDataTypeForClippedSwigluGrad");
70+ const ge::DataType dtype = context->GetInputDataType(1); // x is index 1
71+ ge::graphStatus ret = context->SetOutputDataType(0, dtype);
72+ OP_LOGD(context->GetNodeName(), "End to do InferDataTypeForClippedSwigluGrad");
73+ return ret;
74+}
75+ 
76+IMPL_OP_INFERSHAPE(ClippedSwigluGrad)
77+ .InferShape(InferShapeForClippedSwigluGrad)
78+ .InferDataType(InferDataTypeForClippedSwigluGrad);
79+} // namespace ops
@@ -0,0 +1,366 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+ 
11+/*!
12+ * \file clipped_swiglu_grad_tiling.cpp
13+ * \brief
14+ */
15+#include "clipped_swiglu_grad_tiling.h"
16+#include "../op_kernel/clipped_swiglu_grad_tiling_key.h"
17+ 
18+using Ops::NN::Optiling::TilingRegistry;
19+using namespace ge;
20+namespace optiling {
21+constexpr int64_t Y_GRAD_INDEX = 0;
22+constexpr int64_t X_INDEX = 1;
23+constexpr int64_t GROUP_INDEX_INDEX = 2;
24+constexpr int64_t GRAD_X_INDEX = 0;
25+constexpr int64_t DIM_INDEX = 0;
26+constexpr int64_t ALPHA_INDEX = 1;
27+constexpr int64_t LIMIT_INDEX = 2;
28+constexpr int64_t BIAS_INDEX = 3;
29+constexpr int64_t INTERLEAVED_INDEX = 4;
30+ 
31+constexpr int64_t BLOCK_SIZE = 32;
32+constexpr int64_t BITS_PER_BYTE = 8;
33+constexpr int64_t SWI_FACTOR = 2;
34+constexpr int64_t UB_RESERVE = 1024;
35+constexpr int64_t DB_BUFFER = 2;
36+constexpr int64_t SIZE_OF_FP32 = sizeof(float);
37+ 
38+static int64_t AlignUp(int64_t value, int64_t align) { return (value + align - 1) / align * align; }
39+ 
40+constexpr float DEFAULT_LIMIT = 7.0;
41+constexpr float DEFAULT_ALPHA = 1.702;
42+constexpr float DEFAULT_BIAS = 1.0;
43+ 
44+static const std::set<ge::DataType> SUPPORT_DTYPE = {ge::DT_FLOAT, ge::DT_FLOAT16, ge::DT_BF16};
45+ 
46+ge::graphStatus ClippedSwigluGradTiling::GetShapeAttrsInfo()
47+{
48+ OP_CHECK_IF(CheckAndGetXAndAttrs() != ge::GRAPH_SUCCESS,
49+ OP_LOGE(context_->GetNodeName(), "check x and attrs failed."), return ge::GRAPH_FAILED);
50+ OP_CHECK_IF(CheckAndGetGroupIndex() != ge::GRAPH_SUCCESS,
51+ OP_LOGE(context_->GetNodeName(), "check group_index param failed."), return ge::GRAPH_FAILED);
52+ OP_CHECK_IF(CheckGradX() != ge::GRAPH_SUCCESS, OP_LOGE(context_->GetNodeName(), "check grad_x param failed."),
53+ return ge::GRAPH_FAILED);
54+ return ge::GRAPH_SUCCESS;
55+}
56+ 
57+ge::graphStatus ClippedSwigluGradTiling::GetPlatformInfo()
58+{
59+ auto platformInfo = context_->GetPlatformInfo();
60+ if (platformInfo == nullptr) {
61+ auto compileInfoPtr = context_->GetCompileInfo<ClippedSwigluGradCompileInfo>();
62+ OP_CHECK_IF(compileInfoPtr == nullptr, OP_LOGE(context_, "compile info is null"), return ge::GRAPH_FAILED);
63+ coreNumAll_ = compileInfoPtr->coreNum;
64+ ubSize_ = compileInfoPtr->ubSize;
65+ } else {
66+ auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);
67+ coreNumAll_ = ascendcPlatform.GetCoreNumAiv();
68+ uint64_t ubSizePlatForm;
69+ ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSizePlatForm);
70+ ubSize_ = ubSizePlatForm;
71+ }
72+ return ge::GRAPH_SUCCESS;
73+}
74+ 
75+bool ClippedSwigluGradTiling::IsCapable()
76+{
77+ // 第一版统一实现,所有平台共用一套 tiling,不区分 arch35
78+ return true;
79+}
80+ 
81+ge::graphStatus ClippedSwigluGradTiling::DoOpTiling()
82+{
83+ CountMaxPair();
84+ SetTilingData();
85+ return ge::GRAPH_SUCCESS;
86+}
87+ 
88+void ClippedSwigluGradTiling::SetTilingData()
89+{
90+ tilingData_.set_coreNumAll(coreNumAll_);
91+ tilingData_.set_dimBatchSize(dimBatchSize_);
92+ tilingData_.set_dim2H(dim2H_);
93+ tilingData_.set_isLongH(isLongH_);
94+ tilingData_.set_isGroup(isGroup_);
95+ tilingData_.set_isInterleaved(isInterleaved_);
96+ tilingData_.set_alpha(alpha_);
97+ tilingData_.set_limit(limit_);
98+ tilingData_.set_bias(bias_);
99+ tilingData_.set_ubMaxPair(ubMaxPair_);
100+ tilingData_.set_groupNum(groupNum_);
101+}
102+ 
103+ge::graphStatus ClippedSwigluGradTiling::DoLibApiTiling() { return ge::GRAPH_SUCCESS; }
104+ 
105+ge::graphStatus ClippedSwigluGradTiling::GetWorkspaceSize() { return ge::GRAPH_SUCCESS; }
106+ 
107+ge::graphStatus ClippedSwigluGradTiling::PostTiling()
108+{
109+ context_->SetTilingKey(GetTilingKey());
110+ context_->SetBlockDim(coreNumAll_);
111+ tilingData_.SaveToBuffer(context_->GetRawTilingData()->GetData(), context_->GetRawTilingData()->GetCapacity());
112+ context_->GetRawTilingData()->SetDataSize(tilingData_.GetDataSize());
113+ return ge::GRAPH_SUCCESS;
114+}
115+ 
116+uint64_t ClippedSwigluGradTiling::GetTilingKey() const
117+{
118+ using namespace ClippedSwigluGradArch35Op;
119+ uint64_t isInterleavedKey = (isInterleaved_ != 0) ? TPL_GRAD_INTERLEAVED_TRUE : TPL_GRAD_INTERLEAVED_FALSE;
120+ uint64_t isGroupKey = (isGroup_ != 0) ? TPL_GRAD_GROUP_INDEX : TPL_GRAD_NO_GROUP_INDEX;
121+ return GET_TPL_TILING_KEY(isInterleavedKey, isGroupKey);
122+}
123+ 
124+void ClippedSwigluGradTiling::DumpTilingInfo()
125+{
126+ std::ostringstream info;
127+ info << ", coreNumAll: " << tilingData_.get_coreNumAll();
128+ info << ", ubSize_: " << ubSize_;
129+ info << ", xDims_: " << xDims_;
130+ info << ", cutDim_: " << cutDim_;
131+ info << ", dimBatchSize: " << tilingData_.get_dimBatchSize();
132+ info << ", dim2H: " << tilingData_.get_dim2H();
133+ info << ", isLongH: " << tilingData_.get_isLongH();
134+ info << ", xCutDimNum_: " << xCutDimNum_;
135+ info << ", isGroup: " << tilingData_.get_isGroup();
136+ info << ", isInterleaved: " << tilingData_.get_isInterleaved();
137+ info << ", limit: " << tilingData_.get_limit();
138+ info << ", alpha: " << tilingData_.get_alpha();
139+ info << ", bias: " << tilingData_.get_bias();
140+ info << ", ubMaxPair: " << tilingData_.get_ubMaxPair();
141+ info << ", groupNum: " << tilingData_.get_groupNum();
142+ OP_LOGI(context_->GetNodeName(), "%s", info.str().c_str());
143+}
144+ 
145+ge::graphStatus ClippedSwigluGradTiling::CheckAndGetXAndAttrs()
146+{
147+ auto* attrs = context_->GetAttrs();
148+ OP_CHECK_NULL_WITH_CONTEXT(context_, attrs);
149+ auto* attrDim = attrs->GetAttrPointer<int>(DIM_INDEX);
150+ cutDim_ = attrDim == nullptr ? -1 : *attrDim;
151+ auto* attrAlpha = attrs->GetAttrPointer<float>(ALPHA_INDEX);
152+ alpha_ = attrAlpha == nullptr ? DEFAULT_ALPHA : *attrAlpha;
153+ auto* attrLimit = attrs->GetAttrPointer<float>(LIMIT_INDEX);
154+ limit_ = attrLimit == nullptr ? DEFAULT_LIMIT : *attrLimit;
155+ auto* attrBias = attrs->GetAttrPointer<float>(BIAS_INDEX);
156+ bias_ = attrBias == nullptr ? DEFAULT_BIAS : *attrBias;
157+ auto* attrInterleaved = attrs->GetAttrPointer<bool>(INTERLEAVED_INDEX);
158+ bool interleaved = attrInterleaved == nullptr ? true : *attrInterleaved;
159+ isInterleaved_ = interleaved ? 1 : 0;
160+ OP_CHECK_IF((limit_ <= 0.0f), OP_LOGE(context_->GetNodeName(), "limit must be > 0, but got %f.", limit_),
161+ return ge::GRAPH_FAILED);
162+ auto shapeX = context_->GetInputShape(X_INDEX);
163+ OP_CHECK_NULL_WITH_CONTEXT(context_, shapeX);
164+ const gert::Shape& inputShapeX = shapeX->GetStorageShape();
165+ xDims_ = inputShapeX.GetDimNum();
166+ OP_CHECK_IF(
167+ (cutDim_ > (xDims_ - 1) || cutDim_ < -1 * xDims_),
168+ OP_LOGE(context_->GetNodeName(), "dim should in [-%ld, %ld], but get %ld,", xDims_, xDims_ - 1, cutDim_),
169+ return ge::GRAPH_FAILED);
170+ cutDim_ = cutDim_ < 0 ? (cutDim_ + xDims_) : cutDim_;
171+ xCutDimNum_ = inputShapeX.GetDim(cutDim_);
172+ if (xDims_ == 1) {
173+ dimBatchSize_ = 1;
174+ dim2H_ = inputShapeX.GetDim(0);
175+ } else {
176+ for (int64_t i = 0; i < cutDim_; i++) {
177+ dimBatchSize_ *= inputShapeX.GetDim(i);
178+ }
179+ for (int64_t j = cutDim_; j < xDims_; j++) {
180+ dim2H_ *= inputShapeX.GetDim(j);
181+ }
182+ }
183+ OP_CHECK_IF(
184+ (inputShapeX.GetDim(cutDim_) % 2 == 1),
185+ OP_LOGE(context_->GetNodeName(), "x[dim] should be divisible by 2, but get %ld", inputShapeX.GetDim(cutDim_)),
186+ return ge::GRAPH_FAILED);
187+ auto descX = context_->GetInputDesc(X_INDEX);
188+ OP_CHECK_NULL_WITH_CONTEXT(context_, descX);
189+ xDtype_ = descX->GetDataType();
190+ OP_CHECK_IF((SUPPORT_DTYPE.find(xDtype_) == SUPPORT_DTYPE.end()),
191+ OP_LOGE(context_->GetNodeName(), "x dtype only support float, half or bfloat16, please check."),
192+ return ge::GRAPH_FAILED);
193+ auto shapeGradY = context_->GetInputShape(Y_GRAD_INDEX);
194+ OP_CHECK_NULL_WITH_CONTEXT(context_, shapeGradY);
195+ const gert::Shape& inputShapeGradY = shapeGradY->GetStorageShape();
196+ OP_CHECK_IF((inputShapeGradY.GetDimNum() != xDims_),
197+ OP_LOGE(context_->GetNodeName(), "grad_y rank (%ld) should be equal to x rank (%ld).",
198+ inputShapeGradY.GetDimNum(), xDims_),
199+ return ge::GRAPH_FAILED);
200+ for (int64_t i = 0; i < xDims_; i++) {
201+ if (i == cutDim_) {
202+ OP_CHECK_IF((inputShapeGradY.GetDim(i) != inputShapeX.GetDim(i) / 2),
203+ OP_LOGE(context_->GetNodeName(), "grad_y dim %ld (%ld) should be half of x dim %ld (%ld).", i,
204+ inputShapeGradY.GetDim(i), i, inputShapeX.GetDim(i)),
205+ return ge::GRAPH_FAILED);
206+ } else {
207+ OP_CHECK_IF((inputShapeGradY.GetDim(i) != inputShapeX.GetDim(i)),
208+ OP_LOGE(context_->GetNodeName(), "grad_y dim %ld (%ld) should be equal to x dim %ld (%ld).", i,
209+ inputShapeGradY.GetDim(i), i, inputShapeX.GetDim(i)),
210+ return ge::GRAPH_FAILED);
211+ }
212+ }
213+ auto descGradY = context_->GetInputDesc(Y_GRAD_INDEX);
214+ OP_CHECK_NULL_WITH_CONTEXT(context_, descGradY);
215+ OP_CHECK_IF((descGradY->GetDataType() != xDtype_),
216+ OP_LOGE(context_->GetNodeName(), "grad_y dtype should be the same as x, please check."),
217+ return ge::GRAPH_FAILED);
218+ return ge::GRAPH_SUCCESS;
219+}
220+ 
221+ge::graphStatus ClippedSwigluGradTiling::CheckAndGetGroupIndex()
222+{
223+ auto shapeGroupIndex = context_->GetInputShape(GROUP_INDEX_INDEX);
224+ if (shapeGroupIndex == nullptr) {
225+ isGroup_ = 0;
226+ } else {
227+ const gert::Shape& inputShapeGroupIndex = shapeGroupIndex->GetStorageShape();
228+ int64_t groupIndexDim = inputShapeGroupIndex.GetDimNum();
229+ auto descGroupIndex = context_->GetInputDesc(GROUP_INDEX_INDEX);
230+ OP_CHECK_NULL_WITH_CONTEXT(context_, descGroupIndex);
231+ auto groupIndexDtype = descGroupIndex->GetDataType();
232+ groupNum_ = inputShapeGroupIndex.GetDim(0);
233+ OP_LOGI(context_->GetNodeName(), "group_index: dim=%ld, groupNum=%ld, isNull=%d", groupIndexDim, groupNum_,
234+ shapeGroupIndex == nullptr);
235+ OP_CHECK_IF((groupIndexDim != 1),
236+ OP_LOGE(context_->GetNodeName(),
237+ "the number of dimensions of group_index should be 1, but get %ld.", groupIndexDim),
238+ return ge::GRAPH_FAILED);
239+ OP_CHECK_IF((groupIndexDtype != ge::DT_INT64),
240+ OP_LOGE(context_->GetNodeName(), "groupIndex dtype only support int64, please check."),
241+ return ge::GRAPH_FAILED);
242+ if (groupNum_ == 0) {
243+ isGroup_ = 0;
244+ } else {
245+ isGroup_ = 1;
246+ }
247+ }
248+ return ge::GRAPH_SUCCESS;
249+}
250+ 
251+ge::graphStatus ClippedSwigluGradTiling::CheckGradX()
252+{
253+ auto shapeGradX = context_->GetOutputShape(GRAD_X_INDEX);
254+ OP_CHECK_NULL_WITH_CONTEXT(context_, shapeGradX);
255+ const gert::Shape& outputShapeGradX = shapeGradX->GetStorageShape();
256+ int64_t gradXDims = outputShapeGradX.GetDimNum();
257+ auto descGradX = context_->GetInputDesc(GRAD_X_INDEX);
258+ OP_CHECK_NULL_WITH_CONTEXT(context_, descGradX);
259+ auto gradXDtype = descGradX->GetDataType();
260+ OP_CHECK_IF(
261+ (gradXDims != xDims_),
262+ OP_LOGE(context_->GetNodeName(),
263+ "the number of dimensions of grad_x should be equal to dimensions of x, but get %ld.", gradXDims),
264+ return ge::GRAPH_FAILED);
265+ OP_CHECK_IF((gradXDtype != xDtype_),
266+ OP_LOGE(context_->GetNodeName(), "grad_x dtype should be the same as x, please cheack."),
267+ return ge::GRAPH_FAILED);
268+ return ge::GRAPH_SUCCESS;
269+}
270+ 
271+ge::graphStatus ClippedSwigluGradTiling::CountMaxPair()
272+{
273+ ubMaxPair_ = 1;
274+ // UB 预算系数(单位:sizeof(float)=4 bytes per pair)
275+ // xBuffer: x (双缓冲, A+B两半) = DB*SWI = 4
276+ // dyBuffer: dy (双缓冲, 半宽) = DB = 2
277+ // dxBuffer: dx (单缓冲, 全宽) = SWI = 2
278+ // tmpBuffer1/2: 中间计算量 (A+B两半) = SWI = 2
279+ int64_t xBuffer = DB_BUFFER * SWI_FACTOR;
280+ int64_t dyBuffer = DB_BUFFER;
281+ int64_t dxBuffer = SWI_FACTOR;
282+ int64_t tmpBuffer1 = SWI_FACTOR;
283+ int64_t tmpBuffer2 = SWI_FACTOR;
284+ int64_t totalCoeff = xBuffer + dyBuffer + dxBuffer + tmpBuffer1 + tmpBuffer2;
285+ 
286+ // group_index 直接从 GM GetValue 读取,不占用 UB
287+ int64_t numeratorNoMask = static_cast<int64_t>(ubSize_) - UB_RESERVE;
288+ int64_t prelimPair = ((numeratorNoMask / totalCoeff / BLOCK_SIZE * BLOCK_SIZE) - (BLOCK_SIZE - 1)) / SIZE_OF_FP32;
289+ 
290+ // 第二步:用初步值算 mask buffer 大小
291+ // CompareScalar 输出为 bitmask,1 bit/element,2 个 mask,对齐到 32B
292+ int64_t maskBytes = (prelimPair + BITS_PER_BYTE - 1) / BITS_PER_BYTE;
293+ int64_t maskBufSize = AlignUp(maskBytes, BLOCK_SIZE);
294+ if (maskBufSize < BLOCK_SIZE) {
295+ maskBufSize = BLOCK_SIZE;
296+ }
297+ int64_t maskBuffer = SWI_FACTOR * maskBufSize;
298+ 
299+ // 第三步:减去 mask buffer,重新计算最终 ubMaxPair
300+ int64_t numerator = numeratorNoMask - maskBuffer;
301+ ubMaxPair_ = ((numerator / totalCoeff / BLOCK_SIZE * BLOCK_SIZE) - (BLOCK_SIZE - 1)) / SIZE_OF_FP32;
302+ 
303+ // 对齐 ubMaxPair_ 到 64(CompareScalar 的 CMP_ALIGN),避免 alignedCount 越界
304+ ubMaxPair_ = ubMaxPair_ / 64 * 64;
305+ 
306+ // interleaved 路径限制 ubMaxPair_,避免 SetValue 标量循环过多导致精度问题
307+ // interleaved 路径限制 ubMaxPair_,避免 SetValue 标量循环过多导致精度问题
308+ if (isInterleaved_ == 1) {
309+ int64_t ilMaxPair = 2000;
310+ if (ubMaxPair_ > ilMaxPair) {
311+ ubMaxPair_ = ilMaxPair;
312+ }
313+ }
314+ 
315+ OP_CHECK_IF((numerator <= 0 || ubMaxPair_ <= 0),
316+ OP_LOGE(context_->GetNodeName(), "Input not supported, groupNum is too large."),
317+ return ge::GRAPH_FAILED);
318+ isLongH_ = ubMaxPair_ * SWI_FACTOR < dim2H_ ? 1 : 0;
319+ if (isInterleaved_ == 0 && isLongH_ == 0) {
320+ int64_t sizeofT = (xDtype_ == ge::DT_FLOAT) ? 4 : 2;
321+ int64_t dimH = dim2H_ / SWI_FACTOR;
322+ if ((dimH * sizeofT) % BLOCK_SIZE != 0) {
323+ isLongH_ = 1;
324+ }
325+ }
326+ return ge::GRAPH_SUCCESS;
327+}
328+ 
329+REGISTER_TILING_TEMPLATE("ClippedSwigluGrad", ClippedSwigluGradTiling, 20000);
330+ 
331+ge::graphStatus TilingForClippedSwigluGrad(gert::TilingContext* context)
332+{
333+ // 第一版统一实现,所有平台共用一套 tiling,不区分 arch35
334+ return TilingRegistry::GetInstance().DoTilingImpl(context);
335+}
336+ 
337+ge::graphStatus TilingPrepareForClippedSwigluGrad(gert::TilingParseContext* context)
338+{
339+ OP_LOGD(context, "TilingPrepareForClippedSwigluGrad enter.");
340+ auto compileInfo = context->GetCompiledInfo<ClippedSwigluGradCompileInfo>();
341+ OP_CHECK_NULL_WITH_CONTEXT(context, compileInfo);
342+ auto platformInfo = context->GetPlatformInfo();
343+ OP_CHECK_NULL_WITH_CONTEXT(context, platformInfo);
344+ auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);
345+ compileInfo->coreNum = ascendcPlatform.GetCoreNumAiv();
346+ OP_CHECK_IF((compileInfo->coreNum <= 0),
347+ OP_LOGE(context->GetNodeName(), "Get core num failed, core num: %u",
348+ static_cast<uint32_t>(compileInfo->coreNum)),
349+ return ge::GRAPH_FAILED);
350+ 
351+ uint64_t ubSize;
352+ ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSize);
353+ compileInfo->ubSize = ubSize;
354+ OP_CHECK_IF(
355+ (compileInfo->ubSize <= 0),
356+ OP_LOGE(context->GetNodeName(), "Get ub size failed, ub size: %u", static_cast<uint32_t>(compileInfo->ubSize)),
357+ return ge::GRAPH_FAILED);
358+ 
359+ OP_LOGD(context, "TilingPrepareForClippedSwigluGrad exit.");
360+ return ge::GRAPH_SUCCESS;
361+}
362+ 
363+IMPL_OP_OPTILING(ClippedSwigluGrad)
364+ .Tiling(TilingForClippedSwigluGrad)
365+ .TilingParse<ClippedSwigluGradCompileInfo>(TilingPrepareForClippedSwigluGrad);
366+} // namespace optiling
@@ -0,0 +1,95 @@
1+/**
2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4+ * CANN Open Software License Agreement Version 2.0 (the "License").
5+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+ 
11+/*!
12+ * \file clipped_swiglu_grad_tiling.h
13+ * \brief
14+ */
15+#ifndef AIR_CXX_RUNTIME_V2_OP_IMPL_CLIPPED_SWIGLU_GRAD_H_
16+#define AIR_CXX_RUNTIME_V2_OP_IMPL_CLIPPED_SWIGLU_GRAD_H_
17+ 
18+#include "register/op_impl_registry.h"
19+#include "util/math_util.h"
20+#include "log/log.h"
21+#include "tiling/platform/platform_ascendc.h"
22+#include "platform/platform_infos_def.h"
23+#include "register/tilingdata_base.h"
24+#include "tiling/tiling_api.h"
25+#include "../op_graph/clipped_swiglu_grad_proto.h"
26+#include "op_host/tiling_base.h"
27+#include "op_host/tiling_templates_registry.h"
28+#include "op_host/tiling_util.h"
29+ 
30+namespace optiling {
31+using Ops::NN::Optiling::TilingBaseClass;
32+BEGIN_TILING_DATA_DEF(ClippedSwigluGradTilingData)
33+TILING_DATA_FIELD_DEF(int64_t, coreNumAll);
34+TILING_DATA_FIELD_DEF(int64_t, dimBatchSize);
35+TILING_DATA_FIELD_DEF(int64_t, dim2H);
36+TILING_DATA_FIELD_DEF(int64_t, isLongH);
37+TILING_DATA_FIELD_DEF(int64_t, isGroup);
38+TILING_DATA_FIELD_DEF(int64_t, isInterleaved);
39+TILING_DATA_FIELD_DEF(float, alpha);
40+TILING_DATA_FIELD_DEF(float, limit);
41+TILING_DATA_FIELD_DEF(float, bias);
42+TILING_DATA_FIELD_DEF(int64_t, ubMaxPair);
43+TILING_DATA_FIELD_DEF(int64_t, groupNum);
44+END_TILING_DATA_DEF;
45+ 
46+REGISTER_TILING_DATA_CLASS(ClippedSwigluGrad, ClippedSwigluGradTilingData)
47+ 
48+struct ClippedSwigluGradCompileInfo {
49+ uint64_t coreNum = 0;
50+ uint64_t ubSize = 0;
51+};
52+ 
53+class ClippedSwigluGradTiling : public TilingBaseClass {
54+public:
55+ explicit ClippedSwigluGradTiling(gert::TilingContext* tilingContext) : TilingBaseClass(tilingContext) {}
56+ ~ClippedSwigluGradTiling() override {}
57+ 
58+protected:
59+ bool IsCapable() override;
60+ ge::graphStatus GetPlatformInfo() override;
61+ ge::graphStatus GetShapeAttrsInfo() override;
62+ ge::graphStatus DoOpTiling() override;
63+ ge::graphStatus DoLibApiTiling() override;
64+ uint64_t GetTilingKey() const override;
65+ ge::graphStatus GetWorkspaceSize() override;
66+ ge::graphStatus PostTiling() override;
67+ void DumpTilingInfo() override;
68+ ge::graphStatus CheckAndGetXAndAttrs();
69+ ge::graphStatus CheckAndGetGroupIndex();
70+ ge::graphStatus CheckGradX();
71+ ge::graphStatus CountMaxPair();
72+ void SetTilingData();
73+ 
74+private:
75+ ClippedSwigluGradTilingData tilingData_;
76+ uint64_t coreNumAll_ = 0;
77+ uint64_t ubSize_ = 0;
78+ int64_t xDims_ = 0;
79+ int64_t cutDim_ = 0;
80+ int64_t dimBatchSize_ = 1;
81+ int64_t dim2H_ = 1;
82+ int64_t isLongH_ = 0;
83+ int64_t xCutDimNum_ = 0;
84+ ge::DataType xDtype_ = ge::DT_FLOAT;
85+ int64_t isGroup_ = 0;
86+ int64_t isInterleaved_ = 1;
87+ float limit_ = 0.0;
88+ float alpha_ = 0.0;
89+ float bias_ = 0.0;
90+ int64_t ubMaxPair_ = 0;
91+ int64_t groupNum_ = 0;
92+};
93+ 
94+} // namespace optiling
95+#endif // AIR_CXX_RUNTIME_V2_OP_IMPL_CLIPPED_SWIGLU_GRAD_H_
@@ -0,0 +1,87 @@
1+{
2+ "op_type": "ClippedSwigluGrad",
3+ "op_list": [
4+ {
5+ "bin_filename": "ClippedSwigluGrad_fp16",
6+ "inputs": [
7+ {
8+ "name": "y_grad",
9+ "index": 0,
10+ "dtype": "float16",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [-2]
14+ },
15+ {
16+ "name": "x",
17+ "index": 1,
18+ "dtype": "float16",
19+ "format": "ND",
20+ "paramType": "required",
21+ "shape": [-2]
22+ },
23+ {
24+ "name": "group_index",
25+ "index": 2,
26+ "dtype": "int64",
27+ "format": "ND",
28+ "paramType": "optional",
29+ "shape": [-2]
30+ }
31+ ],
32+ "outputs": [
33+ {
34+ "name": "x_grad",
35+ "index": 0,
36+ "dtype": "float16",
37+ "format": "ND",
38+ "paramType": "required",
39+ "shape": [-2]
40+ }
41+ ],
42+ "attrs": [
43+ { "name": "dim", "dtype": "int64", "value": -1 },
44+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
45+ { "name": "limit", "dtype": "float32", "value": 7.0 },
46+ { "name": "bias", "dtype": "float32", "value": 1.0 },
47+ { "name": "interleaved", "dtype": "bool", "value": true }
48+ ]
49+ },
50+ {
51+ "bin_filename": "ClippedSwigluGrad_fp32",
52+ "inputs": [
53+ { "name": "y_grad", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] },
54+ { "name": "x", "index": 1, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] },
55+ { "name": "group_index", "index": 2, "dtype": "int64", "format": "ND", "paramType": "optional", "shape": [-2] }
56+ ],
57+ "outputs": [
58+ { "name": "x_grad", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] }
59+ ],
60+ "attrs": [
61+ { "name": "dim", "dtype": "int64", "value": -1 },
62+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
63+ { "name": "limit", "dtype": "float32", "value": 7.0 },
64+ { "name": "bias", "dtype": "float32", "value": 1.0 },
65+ { "name": "interleaved", "dtype": "bool", "value": true }
66+ ]
67+ },
68+ {
69+ "bin_filename": "ClippedSwigluGrad_bf16",
70+ "inputs": [
71+ { "name": "y_grad", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] },
72+ { "name": "x", "index": 1, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] },
73+ { "name": "group_index", "index": 2, "dtype": "int64", "format": "ND", "paramType": "optional", "shape": [-2] }
74+ ],
75+ "outputs": [
76+ { "name": "x_grad", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] }
77+ ],
78+ "attrs": [
79+ { "name": "dim", "dtype": "int64", "value": -1 },
80+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
81+ { "name": "limit", "dtype": "float32", "value": 7.0 },
82+ { "name": "bias", "dtype": "float32", "value": 1.0 },
83+ { "name": "interleaved", "dtype": "bool", "value": true }
84+ ]
85+ }
86+ ]
87+}
@@ -0,0 +1,87 @@
1+{
2+ "op_type": "ClippedSwigluGrad",
3+ "op_list": [
4+ {
5+ "bin_filename": "ClippedSwigluGrad_fp16",
6+ "inputs": [
7+ {
8+ "name": "y_grad",
9+ "index": 0,
10+ "dtype": "float16",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [-2]
14+ },
15+ {
16+ "name": "x",
17+ "index": 1,
18+ "dtype": "float16",
19+ "format": "ND",
20+ "paramType": "required",
21+ "shape": [-2]
22+ },
23+ {
24+ "name": "group_index",
25+ "index": 2,
26+ "dtype": "int64",
27+ "format": "ND",
28+ "paramType": "optional",
29+ "shape": [-2]
30+ }
31+ ],
32+ "outputs": [
33+ {
34+ "name": "x_grad",
35+ "index": 0,
36+ "dtype": "float16",
37+ "format": "ND",
38+ "paramType": "required",
39+ "shape": [-2]
40+ }
41+ ],
42+ "attrs": [
43+ { "name": "dim", "dtype": "int64", "value": -1 },
44+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
45+ { "name": "limit", "dtype": "float32", "value": 7.0 },
46+ { "name": "bias", "dtype": "float32", "value": 1.0 },
47+ { "name": "interleaved", "dtype": "bool", "value": true }
48+ ]
49+ },
50+ {
51+ "bin_filename": "ClippedSwigluGrad_fp32",
52+ "inputs": [
53+ { "name": "y_grad", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] },
54+ { "name": "x", "index": 1, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] },
55+ { "name": "group_index", "index": 2, "dtype": "int64", "format": "ND", "paramType": "optional", "shape": [-2] }
56+ ],
57+ "outputs": [
58+ { "name": "x_grad", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] }
59+ ],
60+ "attrs": [
61+ { "name": "dim", "dtype": "int64", "value": -1 },
62+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
63+ { "name": "limit", "dtype": "float32", "value": 7.0 },
64+ { "name": "bias", "dtype": "float32", "value": 1.0 },
65+ { "name": "interleaved", "dtype": "bool", "value": true }
66+ ]
67+ },
68+ {
69+ "bin_filename": "ClippedSwigluGrad_bf16",
70+ "inputs": [
71+ { "name": "y_grad", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] },
72+ { "name": "x", "index": 1, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] },
73+ { "name": "group_index", "index": 2, "dtype": "int64", "format": "ND", "paramType": "optional", "shape": [-2] }
74+ ],
75+ "outputs": [
76+ { "name": "x_grad", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] }
77+ ],
78+ "attrs": [
79+ { "name": "dim", "dtype": "int64", "value": -1 },
80+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
81+ { "name": "limit", "dtype": "float32", "value": 7.0 },
82+ { "name": "bias", "dtype": "float32", "value": 1.0 },
83+ { "name": "interleaved", "dtype": "bool", "value": true }
84+ ]
85+ }
86+ ]
87+}
@@ -0,0 +1,87 @@
1+{
2+ "op_type": "ClippedSwigluGrad",
3+ "op_list": [
4+ {
5+ "bin_filename": "ClippedSwigluGrad_fp16",
6+ "inputs": [
7+ {
8+ "name": "y_grad",
9+ "index": 0,
10+ "dtype": "float16",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [-2]
14+ },
15+ {
16+ "name": "x",
17+ "index": 1,
18+ "dtype": "float16",
19+ "format": "ND",
20+ "paramType": "required",
21+ "shape": [-2]
22+ },
23+ {
24+ "name": "group_index",
25+ "index": 2,
26+ "dtype": "int64",
27+ "format": "ND",
28+ "paramType": "optional",
29+ "shape": [-2]
30+ }
31+ ],
32+ "outputs": [
33+ {
34+ "name": "x_grad",
35+ "index": 0,
36+ "dtype": "float16",
37+ "format": "ND",
38+ "paramType": "required",
39+ "shape": [-2]
40+ }
41+ ],
42+ "attrs": [
43+ { "name": "dim", "dtype": "int64", "value": -1 },
44+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
45+ { "name": "limit", "dtype": "float32", "value": 7.0 },
46+ { "name": "bias", "dtype": "float32", "value": 1.0 },
47+ { "name": "interleaved", "dtype": "bool", "value": true }
48+ ]
49+ },
50+ {
51+ "bin_filename": "ClippedSwigluGrad_fp32",
52+ "inputs": [
53+ { "name": "y_grad", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] },
54+ { "name": "x", "index": 1, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] },
55+ { "name": "group_index", "index": 2, "dtype": "int64", "format": "ND", "paramType": "optional", "shape": [-2] }
56+ ],
57+ "outputs": [
58+ { "name": "x_grad", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2] }
59+ ],
60+ "attrs": [
61+ { "name": "dim", "dtype": "int64", "value": -1 },
62+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
63+ { "name": "limit", "dtype": "float32", "value": 7.0 },
64+ { "name": "bias", "dtype": "float32", "value": 1.0 },
65+ { "name": "interleaved", "dtype": "bool", "value": true }
66+ ]
67+ },
68+ {
69+ "bin_filename": "ClippedSwigluGrad_bf16",
70+ "inputs": [
71+ { "name": "y_grad", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] },
72+ { "name": "x", "index": 1, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] },
73+ { "name": "group_index", "index": 2, "dtype": "int64", "format": "ND", "paramType": "optional", "shape": [-2] }
74+ ],
75+ "outputs": [
76+ { "name": "x_grad", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2] }
77+ ],
78+ "attrs": [
79+ { "name": "dim", "dtype": "int64", "value": -1 },
80+ { "name": "alpha", "dtype": "float32", "value": 1.702 },
81+ { "name": "limit", "dtype": "float32", "value": 7.0 },
82+ { "name": "bias", "dtype": "float32", "value": 1.0 },
83+ { "name": "interleaved", "dtype": "bool", "value": true }
84+ ]
85+ }
86+ ]
87+}
Rcommon/src/framework/logsoftmax_onnx_plugin.cpp→activation/log_softmax_v2/framework/logsoftmax_onnx_plugin.cpp+44-44
Rtorch_extension/tests/utils/common.py→activation/situ_glu/tests/pytest/test_situ_glu_network.py+64-27
Rindex/scatter_nd_update/tests/st/aclnnScatterNd/executor_aclnnScatterNd.py→activation/situ_glu/tests/st/aclnnSituGlu/executor_aclnnSituGlu.py+75-56File content is under review, please try again later
Rindex/scatter_nd_update/examples/test_aclnn_scatter_nd.cpp→index/scatter_nd/examples/test_aclnn_scatter_nd.cpp+0-0
Rindex/scatter_nd_update/tests/ut/op_api/test_aclnn_scatter_nd.cpp→index/scatter_nd/tests/ut/op_api/test_aclnn_scatter_nd.cpp+1-1
Rcommon/src/framework/unique_onnx_plugin.cpp→index/unique_with_counts_and_sorting/framework/unique_onnx_plugin.cpp+0-0
Rcommon/src/framework/batch_matmul_onnx_plugin.cpp→matmul/batch_mat_mul_v3/framework/batch_matmul_onnx_plugin.cpp+37-37
Rcommon/src/framework/matmul_onnx_plugin.cpp→matmul/mat_mul_v3/framework/matmul_onnx_plugin.cpp+67-67
Rcommon/src/framework/batch_normalization_onnx_plugin.cpp→norm/batch_norm/framework/batch_normalization_onnx_plugin.cpp+64-64
Rindex/index_fill_d/op_host/op_api/index_fill.h→norm/bn_training_update_v2/framework/bn_training_update_v2_tf_plugin.cpp+13-16
Rtorch_extension/tests/utils/__init__.py→norm/bn_training_update_v2/op_graph/CMakeLists.txt+1-2
Rcommon/src/framework/instancenormalization_onnx_plugin.cpp→norm/instance_norm_v3/framework/instancenormalization_onnx_plugin.cpp+47-47
Rcommon/src/framework/adaptive_average_pool2d_onnx_plugin.cpp→pooling/adaptive_avg_pool2d/framework/adaptive_average_pool2d_onnx_plugin.cpp+62-62
Rcommon/src/framework/adaptive_max_pool2d_onnx_plugin.cpp→pooling/adaptive_max_pool2d/framework/adaptive_max_pool2d_onnx_plugin.cpp+62-62
Rindex/scatter_add/tests/ut/op_api/CMakeLists.txt→pooling/avg_pool1_d/op_kernel/CMakeLists.txt+8-5
Rloss/multilabel_margin_loss/op_kernel/multilabel_margin_loss.cpp→pooling/avg_pool1_d/op_kernel/arch35/avg_pool1_d.cpp+13-15
Rindex/scatter_nd_update/op_api/aclnn_scatter_nd.h→pooling/avg_pool1_d/tests/ut/op_kernel/avg_pool1_d_tiling_def.h+19-27
Rcommon/src/framework/ascend_anti_quant_onnx_plugin.cpp→quant/ascend_anti_quant/framework/ascend_anti_quant_onnx_plugin.cpp+76-76
Rcommon/src/framework/ascend_dequant_onnx_plugin.cpp→quant/ascend_dequant/framework/ascend_dequant_onnx_plugin.cpp+76-76
Rcommon/src/framework/ascend_quant_onnx_plugin.cpp→quant/ascend_quant/framework/ascend_quant_onnx_plugin.cpp+98-98
Rcommon/src/framework/int8_quantize_onnx_plugin.cpp→quant/ascend_quant/framework/int8_quantize_onnx_plugin.cpp+73-73
Rrnn/dynamic_rnn/op_host/op_api/aclnn_lstm.cpp→rnn/dynamic_rnn/op_api/aclnn_lstm.cpp+66-0
Rrnn/dynamic_rnn/op_host/op_api/aclnn_lstm.h→rnn/dynamic_rnn/op_api/aclnn_lstm.h+1-1
Rrnn/dynamic_rnn/op_host/op_api/dynamic_rnn.cpp→rnn/dynamic_rnn/op_api/dynamic_rnn.cpp+2-1
Rrnn/dynamic_rnn/op_host/op_api/dynamic_rnn.h→rnn/dynamic_rnn/op_api/dynamic_rnn.h+0-0