conv/extend_conv_transpose/op_kernel/CMakeLists.txt 在 master 提交 d7bb08ba1("修复conv3d_backprop_input_v2多申请内存问题")中被误删,导致 extend_conv_transpose 算子在 ascend950 上缺失 op_kernel 编译配置。
conv/extend_conv_transpose/op_kernel/CMakeLists.txt
该算子内核(op_kernel/extend_conv_transpose.cpp)为 MIX 内核(KERNEL_TYPE_MIX_AIC_1_2,AIC+AIV 1:2 混合),复用 conv3d_backprop_input_v2 arch35 的实现头文件,且依赖特定编译选项(-DFORCE_TILING_CONST_PROPAGATION、--cce-no-dcache-flush、-DENABLE_CV_COMM_VIA_SSBUF=true)及 AUTO_SYNC false 配置。配置缺失后,ascend950 量化场景(A16W8,含 bias/scale 输入)路由到该 MIX 内核时触发 AICore 错误。
op_kernel/extend_conv_transpose.cpp
KERNEL_TYPE_MIX_AIC_1_2
-DFORCE_TILING_CONST_PROPAGATION
--cce-no-dcache-flush
-DENABLE_CV_COMM_VIA_SSBUF=true
AUTO_SYNC false
conv/extend_conv_transpose/op_kernel/
extend_conv_transpose.cpp
CMakeLists.txt
ascend950 量化场景下 extend_conv_transpose 正常执行,无 AICore 错误。
ascend950 量化场景用例执行时报 AICore error(MIX 内核路径)。
修复方案:恢复 conv/extend_conv_transpose/op_kernel/CMakeLists.txt,为 ascend950/mc62 计算单元编译内核源文件:
add_kernel_sources( KERNEL_SRC extend_conv_transpose.cpp COMPUTE_UNITS ascend950 mc62 AUTO_SYNC false OPTIONS "-DFORCE_TILING_CONST_PROPAGATION" "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true" )
相关 issue:#2130(extend_conv_transpose 支持 A16W8)、#6079(A16W8 场景小核模板路由统一) 修复 PR:#11415
Describe the current behavior / 问题描述
conv/extend_conv_transpose/op_kernel/CMakeLists.txt在 master 提交 d7bb08ba1("修复conv3d_backprop_input_v2多申请内存问题")中被误删,导致 extend_conv_transpose 算子在 ascend950 上缺失 op_kernel 编译配置。该算子内核(
op_kernel/extend_conv_transpose.cpp)为 MIX 内核(KERNEL_TYPE_MIX_AIC_1_2,AIC+AIV 1:2 混合),复用 conv3d_backprop_input_v2 arch35 的实现头文件,且依赖特定编译选项(-DFORCE_TILING_CONST_PROPAGATION、--cce-no-dcache-flush、-DENABLE_CV_COMM_VIA_SSBUF=true)及AUTO_SYNC false配置。配置缺失后,ascend950 量化场景(A16W8,含 bias/scale 输入)路由到该 MIX 内核时触发 AICore 错误。Environment Information / 环境信息
Steps to reproduce the issue / 重现步骤
conv/extend_conv_transpose/op_kernel/目录下仅剩extend_conv_transpose.cpp,CMakeLists.txt缺失Describe the expected behavior / 预期结果
ascend950 量化场景下 extend_conv_transpose 正常执行,无 AICore 错误。
Related log / screenshot / 日志 / 截图
ascend950 量化场景用例执行时报 AICore error(MIX 内核路径)。
Special notes for this issue/备注
修复方案:恢复
conv/extend_conv_transpose/op_kernel/CMakeLists.txt,为 ascend950/mc62 计算单元编译内核源文件:add_kernel_sources( KERNEL_SRC extend_conv_transpose.cpp COMPUTE_UNITS ascend950 mc62 AUTO_SYNC false OPTIONS "-DFORCE_TILING_CONST_PROPAGATION" "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true" )相关 issue:#2130(extend_conv_transpose 支持 A16W8)、#6079(A16W8 场景小核模板路由统一)
修复 PR:#11415