[Target] Use LLVM target parser for determining Arm(R) A-Profile Architecture features #16425
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Hi @lhutton1 ! Thanks a lot for picking up the recently introduce llvm reflection (well, kind of) for ARM targets too !
Yes I agree, I was also concerned about, all the mentioned target strings are "non-legit" from llvm point of view.
Addition or subtraction can be done with I salute this PR, nice work @lhutton1 ! |
cbalint13
reviewed
Jan 18, 2024
lhutton1
force-pushed
the
improve-aprofile-target-parser
branch
from
dfcdb3e to
e272cfd
Compare
cbalint13
reviewed
Jan 24, 2024
lhutton1
force-pushed
the
improve-aprofile-target-parser
branch
from
037c930 to
abeb0fa
Compare
|
also cc @kparzysz-quic |
|
friendly ping on this |
…itecture features Currently, target features are determined by a set of fixed checks on the target string. This works well for checking support of a small number of simple features, but it doesn't scale. Some problems include: - There are many non-trivial conditions for which a feature may(not) be available. It is easy to miss these with the current implementation. - The inclusion of some features in a target string can imply other features. For example, "+sve" implies "+neon". This currently isn't taken into account. - The tests in tests/cpp/target/parsers/aprofile_test.c suggest that targets such as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon" are supported target strings. The features will be correctly parsed in TVM, however, they are not valid in LLVM. Therefore, it's possible that TVM and LLVM have different understanding of the features available. This commit uses the more robust LLVM target parser to determine support for the features in TVM. It leverages previous infrastructure added to TVM for obtaining a list of all supported features given an input target, and uses this to check the existance of certain features we're interested in. It should be trivial to grow this list over time. As a result of this change, the problems mentioned above are solved. In the current form, this commit drops support for target strings such as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon". A scan of the codebase suggests this functionality is not in use (only in test cases). Should we feel the need to support them, or have a smoother migration for downstream users of TVM we can add a translator to the parser to convert these into LLVM compatible targets. Change-Id: Ic2bf3b68c8af74025ec388d304bd014624c0c585
`ci-gpu` - Made 'codegen' namespacing more specific
`ci-i386` - Required a more modern version of LLVM for the aprofile tests
`ci-hexagon` - Skipped the aprofile tests if LLVM had not been built with
the correct targets
Change-Id: I792b5994fcea52c74b40e040630db1bbd96ca16c
Change-Id: If41e4fd32947a2acddfe9b0691a0c9ba3245d722
Notably, don't abort when encountering a CPU architecture that's not recognised by LLVM. This can happen when compiling with an older version of LLVM. Instead, output a warning. Also add additional checks in the parser for cases when TVM is not compiled with LLVM support and when LLVM is compiled without support for the necessary architectures. Change-Id: I646cb68cadd5462ee2bd694ba5c22ff7dad8f555
Change-Id: Ibfd0beb6dda00aa2a93cd0b47cf28f045e3fde5c
Change-Id: I0b88ecad2987297c428d0f0ca95db35d828c1672
Change-Id: I98a72a95e2b51f8a4b577dcef15f40e7c28719a2
lhutton1
force-pushed
the
improve-aprofile-target-parser
branch
2 times, most recently
from
2e1ce9e to
d3234bb
Compare
|
@tvm-bot rerun |
Change-Id: Iac24bbe31251ebbafacb410abbb67f1e32c171d6
lhutton1
force-pushed
the
improve-aprofile-target-parser
branch
from
d3234bb to
92c68f2
Compare
|
the duplication of the ci-arm CI jobs still appears to be occurring - trying a force push to see if that helps |
Change-Id: Ic8391063d403cc7fe7e2e7d1b4b3c2d6d3bc3146
|
It seems CI was failing due to a memory leak observed when calling The problem seems to come from The reason for this surfacing only now (after previous successful CI runs), is that #16513 was merged, which means the changes in this PR are now run much more frequently in CI. |
|
|
Thanks for the quick response @cbalint13! I didn't try with llvm18 yet, only llvm17. Calling GetAllLLVMCpuFeatures() and GetAllLLVMTargetArches() should reproduce it, but I'll be able to come up with a more concrete example tomorrow |
I'l test it for llvm18 too, just ping me if you have a concrete sample, until than I try invoking as you said (hope to catch it). |
|
Here is a reproducer: #include "tvm/runtime/registry.h"
#include "tvm/target/target.h"
int main() {
auto pf = tvm::runtime::Registry::Get("target.llvm_get_cpu_archlist");
(*pf)(tvm::Target("llvm"));
}Compile: g++ -std=c++17 -O2 -fPIC -I{TVM_DIR}/include -I{TVM_DIR}/3rdparty/dmlc-core/include -I{TVM_DIR}/tvm/3rdparty/dlpack/include -DDMLC_USE_LOGGING_LIBRARY=\<tvm/runtime/logging.h\> -o mem_leak_exec mem_leak.cpp -L{TVM_BUILD_DIR} -ldl -ltvm -pthreadRun with valgrind: LD_PRELOAD="{TVM_BUILD_DIR}/libtvm.so" valgrind --leak-check=full -v --track-origins=yes ./mem_leak_execOutput: |
Thanks a lot for this, I start to look at it now. |
|
This PR causes spurious error messages to be printed when loading
|
Hi @Lunderberg , The messages looks legit, LLVM10 not support those CPU invocations. Possible enhancements here:
This behaviour was intentionate and introduced here @ PR#15761 Later EDIT:
We have direct LLVM awareness, not needing any hardcoded mappings into static lists (unmaintainable IMHO) |
|
The messages look reasonable, based on the available support, but I think they shouldn't be emitted unless the user is attempting to use the invalid target, or making an explicit query of the target parameters. We should not producing an error message when importing TVM. Regarding changing from a warning to an error, even commenting-out the |
|
I agree on the use of internal architectures being preferable to hard-coded lists. The tags were primarily introduced (IIRC) to handle cases where there is no internal architecture that could be queried, such as GPUs with no readily available table to look up, and with limited availability where the compilation shouldn't require queries to a local GPU. |
|
Thanks for the discussion @Lunderberg @cbalint13. I agree that we shouldn't remove the error message completely. Just thinking out loud - the problem here seems to be that the targets registered in tag.cc are parsed when loading tvm, is it possible to defer parsing of these registered targets to when they are actually used by the user? |
Can give me a script line how to reproduce (beside LLVM10 presence) ?
If turns true, we should take this one check out from I take care of this, if you don't mind this it will be a new PR. |
@cbalint13 If you add a new tag to
Thank you, and making a new PR would be perfect! |
quic-sanirudh
added a commit
to quic-sanirudh/tvm
that referenced
this pull request
Mar 28, 2024
This is just a minor fix where the recent [PR apache#16425](apache#16425) seems to have missed this change for LLVM 18 and above, and so we're running into a compilaion failure.
quic-sanirudh
added a commit
that referenced
this pull request
Mar 30, 2024
thaisacs
pushed a commit
to thaisacs/tvm
that referenced
this pull request
Apr 3, 2024
…itecture features (apache#16425) Currently, target features are determined by a set of fixed checks on the target string. This works well for checking support of a small number of simple features, but it doesn't scale. Some problems include: - There are many non-trivial conditions for which a feature may(not) be available. It is easy to miss these with the current implementation. - The inclusion of some features in a target string can imply other features. For example, "+sve" implies "+neon". This currently isn't taken into account. - The tests in tests/cpp/target/parsers/aprofile_test.c suggest that targets such as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon" are supported target strings. The features will be correctly parsed in TVM, however, they are not valid in LLVM. Therefore, it's possible that TVM and LLVM have different understanding of the features available. This commit uses the more robust LLVM target parser to determine support for the features in TVM. It leverages previous infrastructure added to TVM for obtaining a list of all supported features given an input target, and uses this to check the existance of certain features we're interested in. It should be trivial to grow this list over time. As a result of this change, the problems mentioned above are solved. In the current form, this commit drops support for target strings such as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon". A scan of the codebase suggests this functionality is not in use (only in test cases). Should we feel the need to support them, or have a smoother migration for downstream users of TVM we can add a translator to the parser to convert these into LLVM compatible targets.
thaisacs
pushed a commit
to thaisacs/tvm
that referenced
this pull request
Apr 3, 2024
This is just a minor fix where the recent [PR apache#16425](apache#16425) seems to have missed this change for LLVM 18 and above, and so we're running into a compilaion failure.
|
I am also getting additional errors like this is on a LLVM that was built for rocm (AMD platform). We should not send out an error message during static loading time if ARM target is not used, and only have such error message when we attempt to use tags in aprofile. One possible approach is to conditionally register the related tags based on availability of related function |
lhutton1
added a commit
to lhutton1/tvm
that referenced
this pull request
Apr 17, 2024
…port This commit aims to fix the issue described here: apache#16425 (comment) by conditionally registering the target tags based on the availability of the LLVM AArch64 backend. It's possible to extract the targets LLVM has been compiled for using `llvm-config --targets-built`. Change-Id: I20b608aea9ea554b0c0388ee884621305d2d59b9
tqchen
pushed a commit
that referenced
this pull request
Apr 17, 2024
…port (#16897) This commit aims to fix the issue described here: #16425 (comment) by conditionally registering the target tags based on the availability of the LLVM AArch64 backend. It's possible to extract the targets LLVM has been compiled for using `llvm-config --targets-built`. Change-Id: I20b608aea9ea554b0c0388ee884621305d2d59b9
|
It seems that this pull request may lead to a root@e01d939002c0:~/pr_workspace/main_tvm# python
Python 3.10.14 (main, Apr 6 2024, 18:45:05) [GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tvm
Segmentation fault (core dumped)but works with |
I'll look at this. |
|
Thanks @cbalint13 , I was using prebuilt binary at: 10.0.1: clang+llvm-10.0.1-x86_64-linux-gnu-ubuntu-16.04.tar.xz |
|
build with llvm+16.0.0 works on commit |
|
hi @cbalint13 , with bisect debugging, I found that commit root@e01d939002c0:~/pr_workspace/main_tvm# git bisect good
726a1416497eeca7bfb7dcdbd799d00b33c39f79 is the first bad commit
commit 726a1416497eeca7bfb7dcdbd799d00b33c39f79
Author: Luke Hutton <[email protected]>
Date: Wed Mar 27 15:53:46 2024 +0000
[Target] Use LLVM target parser for determining Arm(R) A-Profile Architecture features (#16425)
Currently, target features are determined by a set of fixed checks on
the target string. This works well for checking support of a small
number of simple features, but it doesn't scale. Some problems include:
- There are many non-trivial conditions for which a feature may(not) be
available. It is easy to miss these with the current implementation.
- The inclusion of some features in a target string can imply other
features. For example, "+sve" implies "+neon". This currently isn't
taken into account.
- The tests in tests/cpp/target/parsers/aprofile_test.c suggest that
targets such as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon"
are supported target strings. The features will be correctly parsed in
TVM, however, they are not valid in LLVM. Therefore, it's possible
that TVM and LLVM have different understanding of the features
available.
This commit uses the more robust LLVM target parser to determine support
for the features in TVM. It leverages previous infrastructure added to
TVM for obtaining a list of all supported features given an input
target, and uses this to check the existance of certain features we're
interested in. It should be trivial to grow this list over time. As a
result of this change, the problems mentioned above are solved.
In the current form, this commit drops support for target strings such
as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon". A scan of the
codebase suggests this functionality is not in use (only in test cases).
Should we feel the need to support them, or have a smoother migration
for downstream users of TVM we can add a translator to the parser to
convert these into LLVM compatible targets.
python/tvm/target/codegen.py | 3 +-
src/target/llvm/llvm_instance.cc | 95 ++++----
src/target/llvm/llvm_instance.h | 13 +-
src/target/llvm/llvm_module.cc | 7 +-
src/target/parsers/aprofile.cc | 88 +++----
tests/cpp/target/parsers/aprofile_test.cc | 263 +++++++++++++--------
.../relay/strategy/test_select_implementation.py | 12 +-
tests/python/target/test_llvm_features_info.py | 24 +-
8 files changed, 282 insertions(+), 223 deletions(-) |
I cannot reproduce the crash on my side. |
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements. See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership. The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied. See the License for the
# specific language governing permissions and limitations
# under the License.
#--------------------------------------------------------------------
# Template custom cmake configuration for compiling
#
# This file is used to override the build options in build.
# If you want to change the configuration, please use the following
# steps. Assume you are on the root directory. First copy the this
# file so that any local changes will be ignored by git
#
# $ mkdir build
# $ cp cmake/config.cmake build
#
# Next modify the according entries, and then compile by
#
# $ cd build
# $ cmake ..
#
# Then build in parallel with 8 threads
#
# $ make -j8
#--------------------------------------------------------------------
#---------------------------------------------
# Backend runtimes.
#---------------------------------------------
# Whether enable CUDA during compile,
#
# Possible values:
# - ON: enable CUDA with cmake's auto search
# - OFF: disable CUDA
# - /path/to/cuda: use specific path to cuda toolkit
set(USE_CUDA OFF)
# Whether to enable NCCL support:
# - ON: enable NCCL with cmake's auto search
# - OFF: disable NCCL
# - /path/to/nccl: use specific path to nccl
set(USE_NCCL OFF)
# Whether enable ROCM runtime
#
# Possible values:
# - ON: enable ROCM with cmake's auto search
# - OFF: disable ROCM
# - /path/to/rocm: use specific path to rocm
set(USE_ROCM OFF)
# Whether to enable RCCL support:
# - ON: enable RCCL with cmake's auto search
# - OFF: disable RCCL
# - /path/to/rccl: use specific path to rccl
set(USE_RCCL OFF)
# Whether enable SDAccel runtime
set(USE_SDACCEL OFF)
# Whether enable Intel FPGA SDK for OpenCL (AOCL) runtime
set(USE_AOCL OFF)
# Whether enable OpenCL runtime
#
# Possible values:
# - ON: enable OpenCL with OpenCL wrapper to remove dependency during build
# time and trigger dynamic search and loading of OpenCL in runtime
# - OFF: disable OpenCL
# - /path/to/opencl-sdk: use specific path to opencl-sdk
set(USE_OPENCL OFF)
# Wheather to allow OPENCL cl_mem access to host
# cl_mem will be allocated with CL_MEM_ALLOC_HOST_PTR
# OpenCLWorkspace->GetHostPtr API returns the host accessible pointer
set(USE_OPENCL_ENABLE_HOST_PTR OFF)
# Whether enable Metal runtime
set(USE_METAL OFF)
# Whether enable Vulkan runtime
#
# Possible values:
# - ON: enable Vulkan with cmake's auto search
# - OFF: disable vulkan
# - /path/to/vulkan-sdk: use specific path to vulkan-sdk
set(USE_VULKAN OFF)
# Whether to use spirv-tools.and SPIRV-Headers from Khronos github or gitlab.
#
# Possible values:
# - OFF: not to use
# - /path/to/install: path to your khronis spirv-tools and SPIRV-Headers installation directory
#
set(USE_KHRONOS_SPIRV OFF)
# whether enable SPIRV_KHR_DOT_PRODUCT
set(USE_SPIRV_KHR_INTEGER_DOT_PRODUCT OFF)
# Whether enable OpenGL runtime
set(USE_OPENGL OFF)
# Whether enable MicroTVM runtime
set(USE_MICRO OFF)
# Whether enable RPC runtime
set(USE_RPC ON)
# Whether to build the C++ RPC server binary
set(USE_CPP_RPC OFF)
# Whether to build the C++ native runtime tool binary
set(USE_CPP_RTVM OFF)
# Whether to build the iOS RPC server application
set(USE_IOS_RPC OFF)
# Whether embed stackvm into the runtime
set(USE_STACKVM_RUNTIME OFF)
# Whether enable tiny embedded graph executor.
set(USE_GRAPH_EXECUTOR ON)
# Whether enable tiny graph executor with CUDA Graph
set(USE_GRAPH_EXECUTOR_CUDA_GRAPH OFF)
# Whether enable pipeline executor.
set(USE_PIPELINE_EXECUTOR OFF)
# Whether to enable the profiler for the graph executor and vm
set(USE_PROFILER ON)
# Whether enable microTVM standalone runtime
set(USE_MICRO_STANDALONE_RUNTIME OFF)
# Whether build with LLVM support
# Requires LLVM version >= 4.0
#
# Possible values:
# - ON: enable llvm with cmake's find search
# - OFF: disable llvm, note this will disable CPU codegen
# which is needed for most cases
# - /path/to/llvm-config: enable specific LLVM when multiple llvm-dev is available.
set(USE_LLVM "/home/msra/cy/clang+llvm-13.0.0-x86_64-linux-gnu-ubuntu-20.04/bin/llvm-config --link-static")
set(HIDE_PRIVATE_SYMBOLS ON)
# Whether use MLIR to help analyze, requires USE_LLVM is enabled
# Possible values: ON/OFF
set(USE_MLIR OFF)
#---------------------------------------------
# Contrib libraries
#---------------------------------------------
# Whether to build with BYODT software emulated posit custom datatype
#
# Possible values:
# - ON: enable BYODT posit, requires setting UNIVERSAL_PATH
# - OFF: disable BYODT posit
#
# set(UNIVERSAL_PATH /path/to/stillwater-universal) for ON
set(USE_BYODT_POSIT OFF)
# Whether use BLAS, choices: openblas, atlas, apple
set(USE_BLAS none)
# Whether to use MKL
# Possible values:
# - ON: Enable MKL
# - /path/to/mkl: mkl root path
# - OFF: Disable MKL
# set(USE_MKL /opt/intel/mkl) for UNIX
# set(USE_MKL ../IntelSWTools/compilers_and_libraries_2018/windows/mkl) for WIN32
# set(USE_MKL <path to venv or site-packages directory>) if using `pip install mkl`
set(USE_MKL OFF)
# Whether use DNNL library, aka Intel OneDNN: https://oneapi-src.github.io/oneDNN
#
# Now matmul/dense/conv2d supported by -libs=dnnl,
# and more OP patterns supported in DNNL codegen(json runtime)
#
# choices:
# - ON: Enable DNNL in BYOC and -libs=dnnl, by default using json runtime in DNNL codegen
# - JSON: same as above.
# - C_SRC: use c source runtime in DNNL codegen
# - path/to/oneDNN：oneDNN root path
# - OFF: Disable DNNL
set(USE_DNNL OFF)
# Whether use Intel AMX instructions.
set(USE_AMX OFF)
# Whether use OpenMP thread pool, choices: gnu, intel
# Note: "gnu" uses gomp library, "intel" uses iomp5 library
set(USE_OPENMP none)
# Whether use contrib.random in runtime
set(USE_RANDOM ON)
# Whether use NNPack
set(USE_NNPACK OFF)
# Possible values:
# - ON: enable tflite with cmake's find search
# - OFF: disable tflite
# - /path/to/libtensorflow-lite.a: use specific path to tensorflow lite library
set(USE_TFLITE OFF)
# /path/to/tensorflow: tensorflow root path when use tflite library
set(USE_TENSORFLOW_PATH none)
# Required for full builds with TFLite. Not needed for runtime with TFLite.
# /path/to/flatbuffers: flatbuffers root path when using tflite library
set(USE_FLATBUFFERS_PATH none)
# Possible values:
# - OFF: disable tflite support for edgetpu
# - /path/to/edgetpu: use specific path to edgetpu library
set(USE_EDGETPU OFF)
# Possible values:
# - ON: enable cuDNN with cmake's auto search in CUDA directory
# - OFF: disable cuDNN
# - /path/to/cudnn: use specific path to cuDNN path
set(USE_CUDNN OFF)
# Whether use cuBLAS
set(USE_CUBLAS OFF)
# Whether use MIOpen
set(USE_MIOPEN OFF)
# Whether use MPS
set(USE_MPS OFF)
# Whether use rocBlas
set(USE_ROCBLAS OFF)
# Whether use contrib sort
set(USE_SORT ON)
# Whether to use Arm Compute Library (ACL) codegen
# We provide 2 separate flags since we cannot build the ACL runtime on x86.
# This is useful for cases where you want to cross-compile a relay graph
# on x86 then run on AArch.
#
# An example of how to use this can be found here: docs/deploy/arm_compute_lib.rst.
#
# USE_ARM_COMPUTE_LIB - Support for compiling a relay graph offloading supported
# operators to Arm Compute Library. OFF/ON
# USE_ARM_COMPUTE_LIB_GRAPH_EXECUTOR - Run Arm Compute Library annotated functions via the ACL
# runtime. OFF/ON/"path/to/ACL"
set(USE_ARM_COMPUTE_LIB OFF)
set(USE_ARM_COMPUTE_LIB_GRAPH_EXECUTOR OFF)
# Whether to build with Arm Ethos-N support
# Possible values:
# - OFF: disable Arm Ethos-N support
# - path/to/arm-ethos-N-stack: use a specific version of the
# Ethos-N driver stack
set(USE_ETHOSN OFF)
# If USE_ETHOSN is enabled, use ETHOSN_HW (ON) if Ethos-N hardware is available on this machine
# otherwise use ETHOSN_HW (OFF) to use the software test infrastructure
set(USE_ETHOSN_HW OFF)
# Whether to build with Arm(R) Ethos(TM)-U NPU codegen support
set(USE_ETHOSU OFF)
# Whether to build with CMSIS-NN external library support.
# See https://github.com/ARM-software/CMSIS_5
set(USE_CMSISNN OFF)
# Whether to build with TensorRT codegen or runtime
# Examples are available here: docs/deploy/tensorrt.rst.
#
# USE_TENSORRT_CODEGEN - Support for compiling a relay graph where supported operators are
# offloaded to TensorRT. OFF/ON
# USE_TENSORRT_RUNTIME - Support for running TensorRT compiled modules, requires presense of
# TensorRT library. OFF/ON/"path/to/TensorRT"
set(USE_TENSORRT_CODEGEN OFF)
set(USE_TENSORRT_RUNTIME OFF)
# Whether use VITIS-AI codegen
set(USE_VITIS_AI OFF)
# Build Verilator codegen and runtime
set(USE_VERILATOR OFF)
#Whether to use CLML codegen
set(USE_CLML OFF)
# USE_CLML_GRAPH_EXECUTOR - CLML SDK PATH or ON or OFF
set(USE_CLML_GRAPH_EXECUTOR OFF)
# Build ANTLR parser for Relay text format
# Possible values:
# - ON: enable ANTLR by searching default locations (cmake find_program for antlr4 and /usr/local for jar)
# - OFF: disable ANTLR
# - /path/to/antlr-*-complete.jar: path to specific ANTLR jar file
set(USE_ANTLR OFF)
# Whether use Relay debug mode
set(USE_RELAY_DEBUG OFF)
# Whether to build fast VTA simulator driver
set(USE_VTA_FSIM OFF)
# Whether to build cycle-accurate VTA simulator driver
set(USE_VTA_TSIM OFF)
# Whether to build VTA FPGA driver (device side only)
set(USE_VTA_FPGA OFF)
# Whether use Thrust
set(USE_THRUST OFF)
# Whether use cuRAND
set(USE_CURAND OFF)
# Whether to build the TensorFlow TVMDSOOp module
set(USE_TF_TVMDSOOP OFF)
# Whether to build the PyTorch custom class module
set(USE_PT_TVMDSOOP OFF)
# Whether to use STL's std::unordered_map or TVM's POD compatible Map
set(USE_FALLBACK_STL_MAP OFF)
# Whether to enable Hexagon support
set(USE_HEXAGON OFF)
set(USE_HEXAGON_SDK /path/to/sdk)
# Whether to build the minimal support android rpc server for Hexagon
set(USE_HEXAGON_RPC OFF)
# Hexagon architecture to target when compiling TVM itself (not the target for
# compiling _by_ TVM). This applies to components like the TVM runtime, but is
# also used to select correct include/library paths from the Hexagon SDK when
# building runtime for Android.
# Valid values are v65, v66, v68, v69, v73.
set(USE_HEXAGON_ARCH "v68")
# Whether to use QHL library
set(USE_HEXAGON_QHL OFF)
# Whether to use ONNX codegen
set(USE_TARGET_ONNX OFF)
# Whether enable BNNS runtime
set(USE_BNNS OFF)
# Whether to build static libtvm_runtime.a, the default is to build the dynamic
# version: libtvm_runtime.so.
#
# The static runtime library needs to be linked into executables with the linker
# option --whole-archive (or its equivalent). The reason is that the TVM registry
# mechanism relies on global constructors being executed at program startup.
# Global constructors alone are not sufficient for the linker to consider a
# library member to be used, and some of such library members (object files) may
# not be included in the final executable. This would make the corresponding
# runtime functions to be unavailable to the program.
set(BUILD_STATIC_RUNTIME OFF)
# Caches the build so that building is faster when switching between branches.
# If you switch branches, build and then encounter a linking error, you may
# need to regenerate the build tree through "make .." (the cache will
# still provide significant speedups).
# Possible values:
# - AUTO: search for path to ccache, disable if not found.
# - ON: enable ccache by searching for the path to ccache, report an error if not found
# - OFF: disable ccache
# - /path/to/ccache: use specific path to ccache
set(USE_CCACHE AUTO)
# Whether to use libbacktrace to supply linenumbers on stack traces.
# Possible values:
# - ON: Find libbacktrace from system paths. Report an error if not found.
# - OFF: Don't use libbacktrace.
# - /path/to/libbacktrace: Looking for the libbacktrace header and static lib from a user-provided path. Report error if not found.
# - COMPILE: Build and link to libbacktrace from 3rdparty/libbacktrace.
# - AUTO:
# - Find libbacktrace from system paths.
# - If not found, fallback to COMPILE on Linux or MacOS, fallback to OFF on Windows or other platforms.
set(USE_LIBBACKTRACE AUTO)
# Whether to install a signal handler to print a backtrace on segfault.
# Need to have USE_LIBBACKTRACE enabled.
set(BACKTRACE_ON_SEGFAULT OFF)
# Whether to enable PAPI support in profiling. PAPI provides access to hardware
# counters while profiling.
# Possible values:
# - ON: enable PAPI support. Will search PKG_CONFIG_PATH for a papi.pc
# - OFF: disable PAPI support.
# - /path/to/folder/containing/: Path to folder containing papi.pc.
set(USE_PAPI OFF)
# Whether to use GoogleTest for C++ unit tests. When enabled, the generated
# build file (e.g. Makefile) will have a target "cpptest".
# Possible values:
# - ON: enable GoogleTest. The package `GTest` will be required for cmake
# to succeed.
# - OFF: disable GoogleTest.
# - AUTO: cmake will attempt to find the GTest package, if found GTest will
# be enabled, otherwise it will be disabled.
# Note that cmake will use `find_package` to find GTest. Please use cmake's
# predefined variables to specify the path to the GTest package if needed.
set(USE_GTEST AUTO)
# Enable using CUTLASS as a BYOC backend
# Need to have USE_CUDA=ON
set(USE_CUTLASS OFF)
# Enable to show a summary of TVM options
set(SUMMARIZE OFF)
# Whether to use LibTorch as backend
# To enable pass the path to the root libtorch (or PyTorch) directory
# OFF or /path/to/torch/
set(USE_LIBTORCH OFF)
# Whether to use the Universal Modular Accelerator Interface
set(USE_UMA OFF)
# Set custom Alloc Alignment for device allocated memory ndarray points to
set(USE_KALLOC_ALIGNMENT 64)
# set(USE_LLVM /root/clang+llvm-10.0.1-x86_64-linux-gnu-ubuntu-18.04/bin/llvm-config)
set(USE_LLVM /root/clang+llvm-16.0.0-x86_64-linux-gnu-ubuntu-18.04/bin/llvm-config)
set(USE_CUDA /usr/local/cuda)my config is quite simple, just enabled CUDA and LLVM. It's weird as I can both reproduce this issue on my nvidia-4090 and amd-mi250. my reproduce script is: git checkout 726a1416497eeca7bfb7dcdbd799d00b33c39f79
git submodule update --init --recursive
cd build
cp ../cmake/config.cmake
echo "set(USE_LLVM /root/clang+llvm-16.0.0-x86_64-linux-gnu-ubuntu-18.04/bin/llvm-config)" >> config.cmake
echo "set(USE_CUDA /usr/local/cuda)" >> config.cmake
cmake ..
make -j
cd ..
python -c "import tvm" |
|
Thanks @cbalint13 , BT FULL Trace: gdb.txt |
Still not able to reproduce :-(
|
|
@cbalint13 , after I transfered llvm into 16.0.6, the bug disappeared, it's interesting, thanks. 16.0.1 also works for me. |
|
I had only 16.0.6 (precompiled) at hand for current tests, thinking that issue might be in tvm side. Thank you for you patience and help ! |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Currently, target features are determined by a set of fixed checks on the target string. This works well for checking support of a small number of simple features, but it doesn't scale. Some problems include:
This commit uses the more robust LLVM target parser to determine support for the features in TVM. It leverages previous infrastructure added to TVM for obtaining a list of all supported features given an input target, and uses this to check the existance of certain features we're interested in. It should be trivial to grow this list over time. As a result of this change, the problems mentioned above are solved.
In the current form, this commit drops support for target strings such as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon". A scan of the codebase suggests this functionality is not in use (only in test cases). Should we feel the need to support them, or have a smoother migration for downstream users of TVM we can add a translator to the parser to convert these into LLVM compatible targets.
cc @Mousius @cbalint13 @ekalda @neildhickey