VeriSilicon
undefined symbol: vxBatchGemmNode at NXP imx8mpevk
Hi,
After I built TIM-VX 1.34.fix on libtim-vx.so, it shows that "undefined symbol: vxBatchGemmNode" when I load libtim-vx.so. Then, use ldd libtim-vx.so to check problem. And all .so files can be found. linux-vdso.so.1 (0x0000ffff9a0a6000) libCLC.so => /usr/lib/libCLC.so (0x0000ffff9954b000) libGAL.so => /usr/lib/libGAL.so (0x0000ffff99384000) libOpenVX.so.1 => /usr/lib/libOpenVX.so.1 (0x0000ffff99131000) libOpenVXU.so => /usr/lib/libOpenVXU.so (0x0000ffff99118000) libVSC.so => /usr/lib/libVSC.so (0x0000ffff9817e000) libArchModelSw.so => /usr/lib/libArchModelSw.so (0x0000ffff98156000) libNNArchPerf.so => /usr/lib/libNNArchPerf.so (0x0000ffff98131000) libstdc++.so.6 => /usr/lib/libstdc++.so.6 (0x0000ffff97f5a000) libm.so.6 => /lib/libm.so.6 (0x0000ffff97eb0000) libgcc_s.so.1 => /lib/libgcc_s.so.1 (0x0000ffff97e8b000) libc.so.6 => /lib/libc.so.6 (0x0000ffff97d16000) libpthread.so.0 => /lib/libpthread.so.0 (0x0000ffff97ce7000) /lib/ld-linux-aarch64.so.1 (0x0000ffff9a074000) libdl.so.2 => /lib/libdl.so.2 (0x0000ffff97cd1000)
I use grep with recursive to find vxBatchGemmNode in /usr/lib/ at NXP imx8mpevk, but get empty. The version is Linux imx8mpevk 5.10.52-lts. At VIM3Pro, vxBatchGemmNode is in libOpenVX.so.1.3.0, X86 is also.
@leokuo0725 ,
Please set TIM_VX_USE_EXTERNAL_OVXLIB=ON and build it again. After use this flag, we will try to find ovxlib from /usr/lib/. if it cannot detect the library, you can set OVXLIB_INC OVXLIB_LIB to the correct directory manually.
@sunshinemyson cmake -DCONFIG=YOCTO -DTIM_VX_USE_EXTERNAL_OVXLIB=ON -DOVXLIB_INC="/home/root/TIM-VX/TIM-VX/src/tim/vx/internal/include" .. -- Using YOCTO Project configuration. src/tim/vx -- Using external OVXLIB from /usr/lib -- OVXLIB include directory: /home/root/TIM-VX/TIM-VX/src/tim/vx/internal/include -- Configuring done -- Generating done -- Build files have been written to: /home/root/TIM-VX/TIM-VX/build
[100%] Linking CXX shared library libtim-vx.so /usr/lib/gcc/aarch64-poky-linux/10.2.0/../../../../aarch64-poky-linux/bin/ld: error: /usr/lib: read: Is a directory collect2: error: ld returned 1 exit status make[2]: *** [src/tim/CMakeFiles/tim-vx.dir/build.make:1079: src/tim/libtim-vx.so] Error 1 make[1]: *** [CMakeFiles/Makefile2:133: src/tim/CMakeFiles/tim-vx.dir/all] Error 2 make: *** [Makefile:149: all] Error 2
I don't have idea about OVXLIB_LIB.
OVXLIB_LIB should be the full path direct to libovxlib.so.
Please don't use internal ovxlib if you enabled external ovxlib. You should be able to find the ovxlib headers/lib from the system such as /usr/include/ and /usr/lib/
https://github.com/VeriSilicon/TIM-VX/issues/228 Please check comments in this issue. It might help you solve your problem.
I had a similar problem with OVXLIB, and I had to revert to a previous version to make the program compile. Is there a way to fix this problem without going back to a historical version?
See https://github.com/OAID/Tengine/issues/1257
A dirty patch to fix this problem by providing empty implementation:
create a new cpp file with source code below
#include <cassert>
#include <cstddef>
#include <cstdint>
extern "C"
{
typedef size_t vx_size;
typedef int32_t vx_int32;
typedef int32_t vx_enum;
typedef struct _vx_node *vx_node;
typedef struct _vx_graph *vx_graph;
typedef struct _vx_tensor_t *vx_tensor;
typedef struct _vx_scalar *vx_scalar;
typedef struct _vx_lut *vx_lut;
typedef struct _vx_node *vx_node;
typedef struct _vx_nn_convolution_3d_params_t
{
vx_int32 padding_w_left; /*!< \brief Number of elements added at each side in the left of w dimension of the input. */
vx_int32 padding_w_right; /*!< \brief Number of elements added at each side in the right of w dimension of the input. */
vx_int32 padding_h_top; /*!< \brief Number of elements added at each side in the top of h dimension of the input. */
vx_int32 padding_h_bottom; /*!< \brief Number of elements added at each side in the bottom of h dimension of the input. */
vx_int32 padding_d_front; /*!< \brief Number of elements added at each side in the front of d dimension of the input. */
vx_int32 padding_d_rear; /*!< \brief Number of elements added at each side in the rear of d dimension of the input. */
vx_int32 stride_w; /*!< \brief skip w jump for down scale. */
vx_int32 stride_h; /*!< \brief skip h jump for down scale. */
vx_int32 stride_d; /*!< \brief skip d jump for down scale. */
vx_int32 dilation_w; /*!< \brief "inflate" the kernel by inserting zeros between the kernel elements in the w direction. The value is the number of zeros to insert.*/
vx_int32 dilation_h; /*!< \brief "inflate" the kernel by inserting zeros between the kernel elements in the h direction. The value is the number of zeros to insert.*/
vx_int32 dilation_d; /*!< \brief "inflate" the kernel by inserting zeros between the kernel elements in the d direction. The value is the number of zeros to insert.*/
vx_enum pad_mode; /*!< \brief A VX_TYPE_ENUM of the <tt> \ref vx_pad_mode_e </tt> enumeration. */
vx_scalar pad_const; /*!< \brief pad const value if setting pad mode to const, the const value is base value, not quantized value. */
vx_enum overflow_policy; /*!< \brief A <tt> VX_TYPE_ENUM</tt> of the <tt> vx_convert_policy_e</tt> enumeration. */
vx_enum rounding_policy; /*!< \brief A <tt> VX_TYPE_ENUM</tt> of the <tt> vx_round_policy_e</tt> enumeration. */
vx_enum down_scale_size_rounding; /*!< \brief Rounding method for calculating output dimensions. See <tt>\ref vx_nn_rounding_type_e</tt> */
vx_int32 depth_multiplier; /*!< \brief depthwise multiplier value, if 0, means convolution, elsewise(>=1), the convolution is depthwiseconvolution. */
}vx_nn_convolution_3d_params_t;
vx_node vxBatchGemmNode(vx_graph graph,
vx_tensor matrix_a,
vx_tensor matrix_b,
vx_tensor matrix_c,
vx_scalar trans_a,
vx_scalar trans_b,
vx_scalar trans_c,
vx_tensor output)
{
assert( false && "vxBatchGemmNode is not yet implemented." );
return NULL;
}
vx_node vxTensorTableLookupLayer(
vx_graph graph,
vx_tensor input,
vx_lut InLut,
vx_lut OutLut,
vx_tensor output)
{
assert( false && "vxTensorTableLookupLayer is not yet implemented." );
return NULL;
}
vx_node vxConv3dLayer(vx_graph graph, vx_tensor inputs, vx_tensor weights, vx_tensor biases, const vx_nn_convolution_3d_params_t *convolution_params, vx_size size_of_convolution_params, vx_tensor outputs)
{
assert( false && "vxConv3dLayer is not yet implemented." );
return NULL;
}
}
This will temperately fix the problem, providing you will not use these interfaces in your code.
@fengwang what SDK you use to compile the TIM-VX? Can you please check if there is the VX_BATCH_GEMM_API_SUPPORT makro is defined and set in any header file? This macro should be undefined (or set to 0) for the driver on i.mx8MP.
@robert-kalmar Thanks for reminding me of this. I was using a relatively old version SDK. I shall upgrade it and come back to you.
@robert-kalmar
I upgrade the SDK to the most recent one, with Linux build version Linux imx8mpevk 5.10.72-lts-5.10.y+ga68e31b63f86 #1 SMP PREEMPT Tue Nov 23 06:02:20 UTC 2021 aarch64 aarch64 aarch64 GNU/Linux
But unfortunately I still have those symbols not found.
Still, I noticed this macro is NOT defined in the header files:
A quick test program
#include <algorithm>
#include <iomanip>
#include <iostream>
#include <tuple>
#include <vector>
#include "tim/vx/context.h"
#include "tim/vx/graph.h"
#include "tim/vx/operation.h"
#include "tim/vx/ops/activations.h"
#include "tim/vx/ops/conv2d.h"
#include "tim/vx/ops/fullyconnected.h"
#include "tim/vx/ops/pool2d.h"
#include "tim/vx/ops/softmax.h"
#include "tim/vx/tensor.h"
int main()
{
#ifdef VX_BATCH_GEMM_API_SUPPORT
std::cout << "This macro is defined " << std::endl;
#else
std::cout << "This macro is not defined " << std::endl;
#endif
return 0;
}
gives output:
This macro is not defined