I am following ASPLOS tutorial on a Minisforum UM790 Pro machine with AMD Ryzen NPU. I was able to successfully set up the linux environment, IPU driver, and all Vitis dependencies.
However, when I tried the MM reference example, the results seem to be wrong. Here follows the output from single_core version. The whole_array version's output does not match with ref design either. Only matrix vector version worked.
rm -rf _build
mkdir -p _build
cd _build && cmake -E env CXXFLAGS="-std=c++23 -ggdb" cmake /home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/.. -D CMAKE_C_COMPILER=gcc-13 -D CMAKE_CXX_COMPILER=g++-13 -DTARGET_NAME=matrixMultiplication -Dsubdir=single_core
CMake Deprecation Warning at CMakeLists.txt:14 (cmake_minimum_required):
Compatibility with CMake < 3.5 will be removed from a future version of
CMake.
Update the VERSION argument <min> value or use a ...<max> suffix to tell
CMake that the project does not need compatibility with older versions.
-- The C compiler identification is GNU 13.1.0
-- The CXX compiler identification is GNU 13.1.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/gcc-13 - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/g++-13 - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found Boost: /usr/lib/x86_64-linux-gnu/cmake/Boost-1.74.0/BoostConfig.cmake (found version "1.74.0")
-- Configuring done (0.2s)
-- Generating done (0.0s)
-- Build files have been written to: /home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build
cd _build && cmake --build . --config Release
gmake[1]: Entering directory '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build'
gmake[2]: Entering directory '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build'
gmake[3]: Entering directory '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build'
gmake[3]: Leaving directory '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build'
gmake[3]: Entering directory '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build'
[ 33%] Building CXX object CMakeFiles/matrixMultiplication.dir/home/user/mlir-aie-test/runtime_lib/test_lib/test_utils.cpp.o
[ 66%] Building CXX object CMakeFiles/matrixMultiplication.dir/single_core/test.cpp.o
[100%] Linking CXX executable matrixMultiplication
gmake[3]: Leaving directory '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build'
[100%] Built target matrixMultiplication
gmake[2]: Leaving directory '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build'
gmake[1]: Leaving directory '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/_build'
cp _build/matrixMultiplication matrixMultiplication.exe
mkdir -p build
python3 /home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/aie2.py -m 64 -k 64 -n 64 -M 256 -K 256 -N 256 > build/aie_256x256x256_64x64x64.mlir
mkdir -p build
cd build && xchesscc_wrapper aie2 -I /tools/Xilinx/Vitis/2023.2/aietools/include -DBIT_WIDTH=8 -DDIM_M=64 -DDIM_K=64 -DDIM_N=64 -c /home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/../../../../aie_kernels/aie2/mm.cc -o mm_64x64x64.o
/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/../../../../aie_kernels/aie2/mm.cc:11:9: warning: '__AIENGINE__' macro redefined [-Wmacro-redefined]
#define __AIENGINE__ 2
^
<command line>:3:9: note: previous definition is here
#define __AIENGINE__ 1
^
In file included from /home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/../../../../aie_kernels/aie2/mm.cc:23:
In file included from /tools/Xilinx/Vitis/2023.2/aietools/include/aie_api/aie.hpp:10185:
In file included from /tools/Xilinx/Vitis/2023.2/aietools/include/aie_api/aie_adf.hpp:75:
In file included from /tools/Xilinx/Vitis/2023.2/aietools/include/aie_api/adf/stream.hpp:54:
In file included from /tools/Xilinx/Vitis/2023.2/aietools/include/adf.h:7:
/tools/Xilinx/Vitis/2023.2/aietools/include/adf/intrinsics.h:28:9: warning: 'REL_WRITE' macro redefined [-Wmacro-redefined]
#define REL_WRITE -1
^
/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/../../../../aie_kernels/aie2/mm.cc:20:9: note: previous definition is here
#define REL_WRITE 0
^
2 warnings generated.
Warning in "": (imprecise line-number, the error occurred somewhere in this function): cannot move keep_with_operand operation `v16acc64 ups_w2c(v16int16, uint6_t, uint1_t, uint1_t, uint2_t, bool &)' <344>, would violate structural limitations [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): ... intended destination [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): cannot move keep_with_operand operation `v16acc64 ups_w2c(v16int16, uint6_t, uint1_t, uint1_t, uint2_t, bool &)' <346>, would violate structural limitations [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): ... intended destination [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): cannot move keep_with_operand operation `v16acc64 ups_w2c(v16int16, uint6_t, uint1_t, uint1_t, uint2_t, bool &)' <348>, would violate structural limitations [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): ... intended destination [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): cannot move keep_with_operand operation `v16acc64 ups_w2c(v16int16, uint6_t, uint1_t, uint1_t, uint2_t, bool &)' <350>, would violate structural limitations [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): ... intended destination [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): cannot move keep_with_operand operation `v16acc64 ups_w2c(v16int16, uint6_t, uint1_t, uint1_t, uint2_t, bool &)' <344>, would violate structural limitations [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): ... intended destination [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): cannot move keep_with_operand operation `v16acc64 ups_w2c(v16int16, uint6_t, uint1_t, uint1_t, uint2_t, bool &)' <346>, would violate structural limitations [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): ... intended destination [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): cannot move keep_with_operand operation `v16acc64 ups_w2c(v16int16, uint6_t, uint1_t, uint1_t, uint2_t, bool &)' <348>, would violate structural limitations [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): ... intended destination [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): cannot move keep_with_operand operation `v16acc64 ups_w2c(v16int16, uint6_t, uint1_t, uint1_t, uint2_t, bool &)' <350>, would violate structural limitations [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): ... intended destination [-Wkeep-with-operand]
Warning in "": (imprecise line-number, the error occurred somewhere in this function): loop found to have 4 iterations, fewer than the explicitly annotated minimum 8 [-Wincorrect-annotation]
Warning: : (loop #8)
Non leaf loop was prepared for pipelining. But the pipelined solutions have not been selected.
Consider removing the chess_prepare_for_pipelining directive as it may improve results
mkdir -p build
cd build && aiecc.py --aie-generate-cdo --no-compile-host --xclbin-name=final_256x256x256_64x64x64.xclbin \
--aie-generate-npu --npu-insts-name=insts_256x256x256_64x64x64.txt ../build/aie_256x256x256_64x64x64.mlir
warning: overriding the module target triple with pdarch-unknown-unknown-elf [-Woverride-module]
1 warning generated.
Warning in "": (imprecise line-number, the error occurred somewhere in this function): loop with essential overflow in loop count computation (number of iterations exceeds internal maximum) [-Wloop-count-overflow]
****** Bootgen v2024.1
**** Build date : Jun 18 2024-22:04:45
** Copyright 1986-2022 Xilinx, Inc. All Rights Reserved.
** Copyright 2022-2024 Advanced Micro Devices, Inc. All Rights Reserved.
[INFO] : Bootimage generated successfully
XRT Build Version: 2.18.0 (HEAD)
Build Date: 2024-07-01 14:52:40
Hash ID: 73fe5440974fc51ccaba6366094e4bfa8151f79a
Creating a default 'in-memory' xclbin image.
Section: 'MEM_TOPOLOGY'(6) was successfully added.
Size : 88 bytes
Format : JSON
File : '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/build/aie_256x256x256_64x64x64.mlir.prj/mem_topology.json'
Section: 'AIE_PARTITION'(32) was successfully added.
Size : 12560 bytes
Format : JSON
File : '/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/build/aie_256x256x256_64x64x64.mlir.prj/aie_partition.json'
Info: Embedded Metadata section is missing project.platform.device.core element, adding it.
Successfully wrote (18589 bytes) to the output file: final_256x256x256_64x64x64.xclbin
Leaving xclbinutil.
AIE Compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:00 0:00:01 2/2 4 Workers
Generating: /home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/build/aie_256x256x256_64x64x64.mlir.prj/aie_cdo_elfs.bin
Generating: /home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/build/aie_256x256x256_64x64x64.mlir.prj/aie_cdo_init.bin
Generating: /home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/build/aie_256x256x256_64x64x64.mlir.prj/aie_cdo_enable.bin
export XRT_HACK_UNSECURE_LOADING_XCLBIN=1 && \
./matrixMultiplication.exe -x build/final_256x256x256_64x64x64.xclbin -i build/insts_256x256x256_64x64x64.txt -k MLIR_AIE -M 256 -K 256 -N 256 -v 2 --warmup 1 --iters 1
Matrix size 256x256x256
Sequence instr count: 278
Loading xclbin: build/final_256x256x256_64x64x64.xclbin
Kernel opcode: MLIR_AIE
Name: MLIR_AIE
Registering xclbin: build/final_256x256x256_64x64x64.xclbin
Getting hardware context.
Getting handle to kernel:MLIR_AIE
Writing data into buffer objects.
A =
3.53 3.33 0.56 1.59 0.50 ... 0.59 1.92 1.36 1.98 2.33
3.55 3.38 3.45 3.70 2.58 ... 0.88 0.92 1.83 1.84 0.10
3.81 1.03 2.38 0.70 2.00 ... 3.34 3.78 3.06 3.69 2.83
0.45 3.11 0.21 0.88 1.32 ... 2.92 0.66 0.50 0.61 0.12
1.48 3.42 0.21 3.56 2.23 ... 2.12 2.22 1.54 1.88 3.33
... ... ... ... ... ... ... ... ... ... ...
3.72 2.95 0.41 0.58 1.48 ... 2.20 2.34 3.14 0.74 2.08
3.75 1.64 2.86 0.95 3.70 ... 3.39 2.27 1.22 1.20 1.34
3.95 0.57 1.95 0.05 0.11 ... 0.74 0.05 2.12 1.55 3.98
2.80 2.81 1.49 1.01 1.28 ... 0.57 0.34 3.39 2.62 2.75
2.50 1.41 0.57 3.41 1.20 ... 3.52 3.27 0.61 3.11 3.42
B =
3.75 1.54 1.69 1.72 3.64 ... 3.06 1.29 1.19 3.28 1.34
1.14 1.06 0.57 0.12 2.27 ... 3.28 3.94 3.89 3.25 1.56
0.07 3.58 2.45 3.48 3.94 ... 1.16 1.06 2.17 2.17 3.30
2.31 0.21 3.45 2.09 2.47 ... 0.78 3.30 2.53 0.82 3.20
3.56 1.68 2.25 2.16 2.59 ... 3.12 3.94 2.97 3.23 2.06
... ... ... ... ... ... ... ... ... ... ...
3.38 3.83 1.18 2.48 3.50 ... 1.02 3.92 1.11 3.06 1.11
1.73 1.41 1.59 3.47 1.08 ... 2.22 1.18 3.33 1.08 3.08
0.62 2.20 0.60 0.57 2.62 ... 2.36 3.34 1.81 0.15 1.60
0.96 0.79 3.92 0.95 1.16 ... 0.06 3.02 1.02 1.78 2.42
2.09 2.58 3.58 1.59 1.55 ... 1.66 3.44 3.34 1.09 3.56
Running Kernel (iteration 0).
Running Kernel (iteration 1).
Verifying against reference matmul ...
[ 64, 6] 1024.00 =!= 916.00
[ 64, 7] 1032.00 =!= 912.00
[ 64, 12] 1040.00 =!= 904.00
[ 64, 19] 1088.00 =!= 956.00
[ 64, 20] 1032.00 =!= 928.00
[ 64, 23] 1128.00 =!= 1000.00
[ 64, 24] 1128.00 =!= 996.00
[ 64, 28] 1080.00 =!= 972.00
[ 64, 30] 1112.00 =!= 984.00
[ 64, 31] 1128.00 =!= 1008.00
[ 64, 33] 1024.00 =!= 912.00
[ 64, 35] 1056.00 =!= 940.00
[ 64, 46] 1112.00 =!= 996.00
[ 64, 47] 988.00 =!= 892.00
[ 64, 49] 1088.00 =!= 980.00
[ 64, 50] 1128.00 =!= 992.00
[ 64, 54] 1136.00 =!= 1012.00
[ 64, 56] 1048.00 =!= 940.00
[ 64, 58] 1112.00 =!= 996.00
[ 64, 64] 1048.00 =!= 936.00
[ 64, 68] 1136.00 =!= 1016.00
[ 64, 74] 1080.00 =!= 968.00
[ 64, 79] 1024.00 =!= 900.00
[ 64, 90] 1072.00 =!= 956.00
[ 64, 92] 1048.00 =!= 948.00
[ 64, 95] 1096.00 =!= 988.00
[ 64, 97] 1048.00 =!= 944.00
[ 64, 99] 1040.00 =!= 928.00
[ 64, 133] 952.00 =!= 852.00
[ 64, 156] 1120.00 =!= 1012.00
[ 64, 163] 1072.00 =!= 944.00
[ 64, 174] 1080.00 =!= 876.00
...and 3407 further errors.
Maximum relative error: 21%
Reference:
1008.00 1016.00 984.00 980.00 956.00 ... 936.00 1080.00 1096.00 1064.00 1040.00
1024.00 948.00 944.00 948.00 904.00 ... 928.00 1048.00 1032.00 1072.00 1004.00
996.00 944.00 940.00 956.00 928.00 ... 872.00 1072.00 1024.00 1032.00 1000.00
940.00 912.00 880.00 912.00 900.00 ... 912.00 1032.00 1032.00 1020.00 956.00
1012.00 984.00 988.00 956.00 952.00 ... 936.00 1088.00 1064.00 1040.00 1056.00
... ... ... ... ... ... ... ... ... ... ...
996.00 940.00 924.00 920.00 896.00 ... 924.00 1020.00 996.00 1032.00 1004.00
1056.00 1008.00 988.00 1016.00 996.00 ... 960.00 1104.00 1160.00 1080.00 1112.00
948.00 980.00 976.00 924.00 924.00 ... 896.00 1040.00 1016.00 1040.00 1056.00
1088.00 1016.00 996.00 1020.00 976.00 ... 960.00 1088.00 1104.00 1120.00 1080.00
1048.00 1020.00 1020.00 1048.00 1024.00 ... 952.00 1152.00 1112.00 1112.00 1112.00
Output:
1004.00 1008.00 980.00 972.00 948.00 ... 932.00 1072.00 1080.00 1056.00 1040.00
1016.00 944.00 936.00 944.00 896.00 ... 924.00 1040.00 1020.00 1064.00 1000.00
988.00 940.00 928.00 952.00 920.00 ... 864.00 1056.00 1020.00 1024.00 996.00
936.00 908.00 872.00 904.00 892.00 ... 904.00 1024.00 1024.00 1012.00 952.00
1004.00 976.00 980.00 952.00 944.00 ... 928.00 1080.00 1056.00 1032.00 1048.00
... ... ... ... ... ... ... ... ... ... ...
1000.00 952.00 916.00 976.00 984.00 ... 900.00 1056.00 1032.00 1040.00 1056.00
928.00 908.00 936.00 916.00 920.00 ... 872.00 1032.00 988.00 996.00 996.00
988.00 968.00 1020.00 996.00 968.00 ... 924.00 1072.00 1048.00 1072.00 1024.00
1056.00 1032.00 1032.00 988.00 1000.00 ... 996.00 1096.00 1104.00 1112.00 1088.00
1056.00 996.00 940.00 988.00 980.00 ... 948.00 1072.00 1096.00 1128.00 1056.00
Verify time: 0.00 s.
Avg NPU matmul time: 834.00us.
Avg NPU gflops: 40.23
Min NPU matmul time: 834.00us.
Max NPU gflops: 40.23
Max NPU matmul time: 834.00us.
Min NPU gflops: 40.23
Error count: 3439
Failed.
make: *** [/home/user/mlir-aie-test/programming_examples/basic/matrix_multiplication/single_core/../makefile-common:87: run] Error 1
makslevental
commented
3 months ago
Hi,
I am following ASPLOS tutorial on a Minisforum UM790 Pro machine with AMD Ryzen NPU. I was able to successfully set up the linux environment, IPU driver, and all Vitis dependencies.
However, when I tried the MM reference example, the results seem to be wrong. Here follows the output from single_core version. The
whole_arrayversion's output does not match with ref design either. Only matrix vector version worked.