rocm-arch / tensorflow-rocm

tensorflow-rocm AUR package
17 stars 12 forks source link

Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+ #14

Closed lubosz closed 4 years ago

lubosz commented 4 years ago

Hi, i'm trying to build tensorflow-rocm and get the following errors during the build:

INFO: From Compiling tensorflow/core/kernels/gather_functor_gpu.cu.cc [for host]:
clang-11: warning: /usr/bin/gcc-9: 'linker' input unused [-Wunused-command-line-argument]
clang-11: warning: argument unused during compilation: '--hip-device-lib-path=/opt/rocm/lib' [-Wunused-command-line-argument]
clang-11: warning: /usr/bin/gcc-9: 'linker' input unused [-Wunused-command-line-argument]
INFO: From Compiling tensorflow/core/kernels/gather_functor_batched_gpu.cu.cc [for host]:
clang-11: warning: /usr/bin/gcc-9: 'linker' input unused [-Wunused-command-line-argument]
clang-11: warning: argument unused during compilation: '--hip-device-lib-path=/opt/rocm/lib' [-Wunused-command-line-argument]
clang-11: warning: /usr/bin/gcc-9: 'linker' input unused [-Wunused-command-line-argument]
INFO: From Compiling tensorflow/core/kernels/dynamic_partition_op_gpu.cu.cc [for host]:
clang-11: warning: /usr/bin/gcc-9: 'linker' input unused [-Wunused-command-line-argument]
clang-11: warning: argument unused during compilation: '--hip-device-lib-path=/opt/rocm/lib' [-Wunused-command-line-argument]
clang-11: warning: /usr/bin/gcc-9: 'linker' input unused [-Wunused-command-line-argument]
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr9 = V_AND_B32_dpp undef $vgpr9(tied-def 0), $vgpr7, killed $vgpr10, 322, 15, 15, 1, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr8 = V_MOV_B32_dpp killed $vgpr8(tied-def 0), $vgpr7, 323, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr6 = V_AND_B32_dpp undef $vgpr6(tied-def 0), $vgpr4, killed $vgpr7, 322, 15, 15, 1, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr5 = V_MOV_B32_dpp killed $vgpr5(tied-def 0), $vgpr4, 323, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr9 = V_AND_B32_dpp undef $vgpr9(tied-def 0), $vgpr7, killed $vgpr10, 322, 15, 15, 1, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr8 = V_MOV_B32_dpp killed $vgpr8(tied-def 0), $vgpr7, 323, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr6 = V_AND_B32_dpp undef $vgpr6(tied-def 0), $vgpr4, killed $vgpr7, 322, 15, 15, 1, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr5 = V_MOV_B32_dpp killed $vgpr5(tied-def 0), $vgpr4, 323, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr10 = V_AND_B32_dpp undef $vgpr10(tied-def 0), $vgpr4, killed $vgpr11, 322, 15, 15, 1, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr19 = V_MOV_B32_dpp killed $vgpr19(tied-def 0), $vgpr14, 323, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr105 = V_AND_B32_dpp undef $vgpr105(tied-def 0), $vgpr35, $vgpr104, 322, 15, 15, 1, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr103 = V_MOV_B32_dpp killed $vgpr103(tied-def 0), $vgpr102, 322, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr104 = V_MOV_B32_dpp killed $vgpr104(tied-def 0), $vgpr101, 323, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr99 = V_MOV_B32_dpp killed $vgpr99(tied-def 0), $vgpr102, 323, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr17 = V_AND_B32_dpp undef $vgpr17(tied-def 0), $vgpr14, $vgpr15, 322, 15, 15, 1, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr19 = V_MOV_B32_dpp killed $vgpr19(tied-def 0), $vgpr13, 322, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr19 = V_MOV_B32_dpp killed $vgpr19(tied-def 0), $vgpr14, 323, 15, 15, 0, implicit $exec
error: Illegal instruction detected: Invalid dpp_ctrl value: broadcasts are not supported on GFX10+
renamable $vgpr12 = V_MOV_B32_dpp killed $vgpr12(tied-def 0), $vgpr13, 323, 15, 15, 0, implicit $exec
clang-11: error: amdgcn-link command failed with exit code 1 (use -v to see invocation)
ERROR: /home/bmonkey/.cache/yay/tensorflow-rocm/src/tensorflow-2.3.0-rocm/tensorflow/core/kernels/BUILD:2677:18: output 'tensorflow/core/kernels/_objs/dynamic_partition_op_gpu/dynamic_partition_op_gpu.cu.pic.o' was not created
ERROR: /home/bmonkey/.cache/yay/tensorflow-rocm/src/tensorflow-2.3.0-rocm/tensorflow/core/kernels/BUILD:2677:18: not all outputs were created or valid
ERROR: /home/bmonkey/.cache/yay/tensorflow-rocm/src/tensorflow-2.3.0-rocm/tensorflow/python/tools/BUILD:99:10 not all outputs were created or valid
INFO: Elapsed time: 512.233s, Critical Path: 133.12s
INFO: 4523 processes: 4523 local.
FAILED: Build did NOT complete successfully
==> ERROR: A failure occurred in build().
    Aborting...
acxz commented 4 years ago

My guess is that gfx10+ is not supported by rocm. See: https://github.com/RadeonOpenCompute/ROCm#Hardware-and-Software-Support

There is not much we can do on the packaging end about this.

lubosz commented 4 years ago

Ok, thanks for the reply. Hopefully support for Navi will arrive before I buy a new GPU.