已关闭
trans_data算子新增ascend350平台支持 #5489
马琦钧创建于 22 天前关闭于 16 天前
trans_data算子新增ascend350平台支持 #5489
已关闭
共 5 个文件变更+1299-3
| @@ -10,9 +10,9 @@ | |||
| 10 | # ---------------------------------------------------------------------------- | 10 | # ---------------------------------------------------------------------------- |
| 11 | 11 | ||
| 12 | # 设置算子定义时支持的芯片类型 | 12 | # 设置算子定义时支持的芯片类型 |
| 13 | -set(SUPPORT_COMPUTE_UNIT "ascend950") | 13 | +set(SUPPORT_COMPUTE_UNIT "ascend950" "ascend350") |
 | |||
| 14 | # 设置每种芯片类型对应的tiling文件目录，即采用op_host目录下哪个文件夹下的tiling文件编译 | 14 | # 设置每种芯片类型对应的tiling文件目录，即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 15 | -set(SUPPORT_TILING_DIR "arch35") | 15 | +set(SUPPORT_TILING_DIR "arch35" "arch35") |
| 16 | add_all_modules_sources(OPTYPE trans_data ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE HOSTCPU TRUE) | 16 | add_all_modules_sources(OPTYPE trans_data ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE HOSTCPU TRUE) |
| 17 | 17 | ||
| 18 | # format_transfer 辅助源文件不匹配 *_aicpu.cpp glob，需显式追加到已创建的 OBJECT 目标 | 18 | # format_transfer 辅助源文件不匹配 *_aicpu.cpp glob，需显式追加到已创建的 OBJECT 目标 |
| @@ -0,0 +1,2 @@ | |||
| 1 | +[TransData] | ||
| 2 | +default=0 | ||
| @@ -126,6 +126,7 @@ public: | |||
| 126 | .ExtendCfgInfo("opFile.value", "trans_data_apt"); | 126 | .ExtendCfgInfo("opFile.value", "trans_data_apt"); |
| 127 | 127 | ||
| 128 | this->AICore().AddConfig("ascend950", aicore_config); | 128 | this->AICore().AddConfig("ascend950", aicore_config); |
| 129 | + this->AICore().AddConfig("ascend350", aicore_config); | ||
| 129 | } | 130 | } |
| 130 | }; | 131 | }; |
| 131 | 132 | ||
| @@ -449,7 +449,7 @@ | |||
| 449 | {"name":"BroadcastTo", "compute_units": ["ascend950", "ascend350"], "auto_sync": false, "impl_mode":"", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"]}}, | 449 | {"name":"BroadcastTo", "compute_units": ["ascend950", "ascend350"], "auto_sync": false, "impl_mode":"", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"]}}, |
| 450 | {"name": "DynamicPartition", "compute_units": ["ascend950", "ascend350"], "auto_sync": false, "impl_mode": "", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"]}}, | 450 | {"name": "DynamicPartition", "compute_units": ["ascend950", "ascend350"], "auto_sync": false, "impl_mode": "", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"]}}, |
| 451 | {"name":"Tile", "compute_units": ["ascend950", "ascend350", "mc62"], "auto_sync": false, "impl_mode":"", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "mc62": ["--cce-no-dcache-flush"]}}, | 451 | {"name":"Tile", "compute_units": ["ascend950", "ascend350", "mc62"], "auto_sync": false, "impl_mode":"", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "mc62": ["--cce-no-dcache-flush"]}}, |
| 452 | - {"name":"TransData", "compute_units": ["ascend950"], "auto_sync": false, "impl_mode": "", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"]}}, | 452 | + {"name":"TransData", "compute_units": ["ascend950", "ascend350"], "auto_sync": false, "impl_mode": "", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"]}}, |
| 453 | {"name":"Expand", "compute_units": ["ascend950", "mc62", "ascend350"], "auto_sync": false, "impl_mode":"", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "mc62": ["--cce-no-dcache-flush"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"]}}, | 453 | {"name":"Expand", "compute_units": ["ascend950", "mc62", "ascend350"], "auto_sync": false, "impl_mode":"", "compile_options": {"ascend950": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"], "mc62": ["--cce-no-dcache-flush"], "ascend350": ["-mllvm -cce-aicore-dcci-before-kernel-end=false"]}}, |
| 454 | {"name":"GeluQuant", "compute_units": ["ascend310p", "ascend910b", "ascend910", "ascend950"], "auto_sync" : true}, | 454 | {"name":"GeluQuant", "compute_units": ["ascend310p", "ascend910b", "ascend910", "ascend950"], "auto_sync" : true}, |
| 455 | {"name":"DequantSwigluQuant", "compute_units": ["ascend910b","ascend910_93", "ascend950"], "auto_sync" : true}, | 455 | {"name":"DequantSwigluQuant", "compute_units": ["ascend910b","ascend910_93", "ascend950"], "auto_sync" : true}, |
1