How to execute a VCK5000-based mlir-tutorial?

[shibizhao]

Hi,
I am new of the mlir-aie. And I have installed the mlir-aie according to the document. And I would like to execute the mlir_tutorials on VCK5000.

First, I check the installed "install/runtime_lib" directory. There is only an "x86_64-hsa".

Secondly, I found the makefile-common is used for aarch64 (I guess it is for VCK190). So I would like to know how I should modify the makefile-common for VCK5000.

Thanks.

[shibizhao]

Update, I tried to modify the makefile-common as:

# Contains common definitions used across the Makefiles of all tutorials.

# MLIR-AIE install directory. If you have sourced utils/env_setup.sh before
# running make, the following should work to find the AIE install directory.
AIE_RUNTIME_LIB ?= $(shell realpath $(dir $(shell which aie-opt))/../runtime_lib)
AIE_INSTALL ?= $(AIE_RUNTIME_LIB)/x86_64-hsa


# VITIS related variables
VITIS_ROOT ?= $(shell realpath $(dir $(shell which vitis))/../)
VITIS_AIETOOLS_DIR ?= ${VITIS_ROOT}/aietools
VITIS_AIE_INCLUDE_DIR ?= ${VITIS_ROOT}/aietools/data/versal_prod/lib
VITIS_AIE2_INCLUDE_DIR ?= ${VITIS_ROOT}/aietools/data/aie_ml/lib


#${VITIS_ROOT}/gnu/aarch64/lin/aarch64-linux/aarch64-xilinx-linux/

# The libstdc++ version that is installed in the sysroot given above. This is
# used for include and library paths. If you built the sysroot with Vitis
# 2022.2 and PetaLinux 2022.2, libstdc++ 11.2.0 will be installed. 
# LIBCXX_VERSION ?= 3.4.30

# The following flags are passed to both AI core and host compilation for
# aiecc.py invocations.
AIECC_FLAGS += --host-target=x86_64-linux-gnu

CHESSCC_FLAGS = -f -p me -P ${VITIS_AIE_INCLUDE_DIR} -I ${VITIS_AIETOOLS_DIR}/include
CHESS_FLAGS = -P ${VITIS_AIE_INCLUDE_DIR}

# The following additional flags are only applied for host code.
AIECC_HOST_FLAGS += \
    -I$(AIE_INSTALL)/test_lib/include \
    -I${AIE_INSTALL}/xaiengine/include \
    --gcc-toolchain=/usr \
    -L$(AIE_INSTALL)/test_lib/lib -ltest_lib 

# Add the
# necessary search paths for the sysroot so clang++ can find the aarch64
# includes and libraries. Some of these shouldn't be necessary, except that
# sysroot shipped with Vitis is slightly broken, so clang can't find things
# automatically just using --gcc-toolchain
AIECC_HOST_FLAGS += \
    -I/usr/include/c++/12 \
    -I/usr/include \
    -I/usr/include/x86_64-linux-gnu/c++/12 \
    -L/usr/lib/x86_64-linux-gnu/ \
    -B/usr/lib/x86_64-linux-gnu/

And I entered to tutorial-1 dir and compiled the elf:

(sandbox) (mlir-aie-dev) bizhao.shi@server:~/research/compiler/mlir-aie/mlir_tutorials_vck5000/tutorial-1$ make core_1_4.elf
aiecc.py -j4 aie.mlir
Found xchesscc at /rshome/software/Xilinx/Vitis/2023.2/aietools
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:01 0/1 1 Workerwarning: overriding the module target triple with pdarch-unknown-unknown-elf [-Woverride-module]
1 warning generated.
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:01 0/1 1 WorkerWarning in "../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_chess.h", line 647, column 131: ignoring attribute [[deprecated]] on class declaration
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:02 0/1 1 WorkerWarning in "../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_common.h", line 63, column 132: ignoring attribute [[deprecated]] on class declaration
Warning in "../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_common.h", line 96, column 131: ignoring attribute [[deprecated]] on class declaration
 AIE Compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:00 0:00:03 2/2 4 Workers

Maybe it works. But when I tried to compile tutorial-1.exe, there was an error:

(sandbox) (mlir-aie-dev) bizhao.shi@server:~/research/compiler/mlir-aie/mlir_tutorials_vck5000/tutorial-1$ make tutorial-1.exe
aiecc.py -j4 --host-target=x86_64-linux-gnu aie.mlir -I/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64-hsa/test_lib/include -I/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64-hsa/xaiengine/include --gcc-toolchain=/usr -L/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64-hsa/test_lib/lib -ltest_lib  -I/usr/include/c++/12 -I/usr/include -I/usr/include/x86_64-linux-gnu/c++/12 -L/usr/lib/x86_64-linux-gnu/ -B/usr/lib/x86_64-linux-gnu/ ./test.cpp -o tutorial-1.exe
Found xchesscc at /rshome/software/Xilinx/Vitis/2023.2/aietools
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:01 0/1 1 Workerwarning: overriding the module target triple with pdarch-unknown-unknown-elf [-Woverride-module]
1 warning generated.
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:01 0/1 1 WorkerWarning in "../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_chess.h", line 647, column 131: ignoring attribute [[deprecated]] on class declaration
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:02 0/1 1 WorkerWarning in "../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_common.h", line 63, column 132: ignoring attribute [[deprecated]] on class declaration
Warning in "../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_common.h", line 96, column 131: ignoring attribute [[deprecated]] on class declaration
 AIE Compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:00 0/2 4 Workersclang: error: no such file or directory: '/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64/test_lib/lib/libmemory_allocator_ion.a'
Error encountered while running: clang++ -std=c++17 --target=x86_64-linux-gnu /rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64/test_lib/lib/libmemory_allocator_ion.a -I/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64/xaiengine/include -L/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64/xaiengine/lib -L/rshome/software/Xilinx/Vitis/2023.2/aietools/lib/lnx64.o -Wl,-R/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64/xaiengine/lib -I/rshome/bizhao.shi/research/compiler/mlir-aie/mlir_tutorials_vck5000/tutorial-1/aie.mlir.prj -fuse-ld=lld -lm -lxaiengine -D__AIEARCH__=10 -I/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64-hsa/test_lib/include -I/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64-hsa/xaiengine/include --gcc-toolchain=/usr -L/rshome/bizhao.shi/research/compiler/mlir-aie/install/runtime_lib/x86_64-hsa/test_lib/lib -ltest_lib -I/usr/include/c++/12 -I/usr/include -I/usr/include/x86_64-linux-gnu/c++/12 -L/usr/lib/x86_64-linux-gnu/ -B/usr/lib/x86_64-linux-gnu/ ./test.cpp -o tutorial-1.exe
 AIE Compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:00 0/2 4 Workers
make: *** [Makefile:24: tutorial-1.exe] Error 1

But there is not a subdirectory named "x86_64" in the runtime_lib.

I am confused.

[eddierichter-amd]

Thanks for your questions! The mlir-tutorials came out before our VCK5000 platform. Are you able to use the programming examples that we provide? The vector vector add example that we have has a Makefile that can compile either on Ryzen AI or VCK5000.

[shibizhao]

Thanks for your reply!!!!

I executed the "make vck5000" and the "./test.elf" commands following the readme in the vector_vector_add example.
It could not run with output:

(sandbox) (mlir-aie-dev) bizhao.shi@server:~/research/compiler/mlir-aie/programming_examples/basic/vector_vector_add$ make vck5000
mkdir -p build
python3 /rshome/bizhao.shi/research/compiler/mlir-aie/programming_examples/basic/vector_vector_add/aie2.py xcvc1902 6 > build/aie.mlir
aiecc.py \
    --link_against_hsa --host-target=x86_64-amd-linux-gnu build/aie.mlir \                                                  -I/rshome/bizhao.shi/research/compiler/mlir-aie/programming_examples/basic/vector_vector_add/../../../install/runtime_lib/x86_64-hsa/test_lib/include \
        /rshome/bizhao.shi/research/compiler/mlir-aie/programming_examples/basic/vector_vector_add/test_vck5000.cpp \
        /rshome/bizhao.shi/research/compiler/mlir-aie/programming_examples/basic/vector_vector_add/../../../install/runtime_lib/x86_64-hsa/test_lib/src/test_library.cpp \                                                                      -Wl,--whole-archive -Wl,--no-whole-archive -lstdc++ -ldl -lelf -o test.elf
Found xchesscc at /rshome/software/Xilinx/Vitis/2023.2/aietools
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:01 0/1 1 Workerwarning: overriding the module target triple with pdarch-unknown-unknown-elf [-Woverride-module]
1 warning generated.
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:01 0/1 1 WorkerWarning in "../../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_chess.h", line 647, column 131: ignoring attribute [[deprecated]] on class declaration
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:02 0/1 1 WorkerWarning in "../../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_common.h", line 63, column 132: ignoring attribute [[deprecated]] on class declaration
Warning in "../../../../../../../software/Xilinx/Vitis/2023.2/aietools/data/versal_prod/lib/me_common.h", line 96, column 131: ignoring attribute [[deprecated]] on class declaration
 MLIR compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:-- 0:00:03 0/1 1 WorkerWarning 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]
 AIE Compilation: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:00 0:00:04 2/2 4 Workers
(sandbox) (mlir-aie-dev) bizhao.shi@server:~/research/compiler/mlir-aie/programming_examples/basic/vector_vector_add$ ./test.elf
hsa_init failed
[ERROR] Error when calling mlir_aie_init_device)

I have sourced the scripts of xrt and vitis 2023.2. And my server has three FPGAs:

(sandbox) (mlir-aie-dev) bizhao.shi@server:~/research/compiler/mlir-aie/programming_examples/basic/vector_vector_add$ xbutil validate
ERROR: Please specify a device using --device option
 Available devices:
  [0000:3b:00.1] : xilinx_u200_gen3x16_xdma_base_2
  [0000:86:00.1] : xilinx_u280_gen3x16_xdma_base_1
  [0000:af:00.1] : xilinx_vck5000_gen4x8_qdma_base_2

So I think maybe it is the default device index in "mlir_aie_init_device" function is incorrect.

I modified test_vck5000.cpp in the vector_vector_add example as:

int main(int argc, char *argv[]) {
  uint64_t row = 0;
  uint64_t col = 6;

  std::vector<hsa_queue_t *> queues;
  uint32_t aie_max_queue_size(0);

  aie_libxaie_ctx_t *xaie = mlir_aie_init_libxaie();

  // This is going to initialize HSA, create a queue
  // and get an agent
 // int ret = mlir_aie_init_device(xaie);
 int ret = mlir_aie_init_device(xaie, 2);

  if (ret) {
    std::cout << "[ERROR] Error when calling mlir_aie_init_device)"
              << std::endl;
    return -1;
  }

The error is the same.

[shibizhao]

Update:

I have install the AIR PCIe kernel and driver for VCK5000.

And now I execute the vector_vector_add in the root mode, there is no any output and (maybe stuck at mlir-aie-init-device).

[eddierichter-amd]

What PDI do you have loaded on the VCK5000? For mlir-aie we use a PDI image that is different than the standard XRT shell. Instructions for how to obtain and load the image is in the platform repo: https://github.com/Xilinx/ROCm-air-platforms/blob/main/platform/vck5000/README.md.

If you do have the ROCm-air-platforms PDI loaded on the card, can you let me know what you see in dmesg when you insert the driver?

[shibizhao]

Hi,

Thanks for your reply.

I have programmed vck5000 using the platform in ROCm-air-platforms.

The driver amdair.ko has been added and used by this vck5000. (from lspci -vd 10ee:).

The dmesg shows as:

[71314.406557] amdair amdair: amdair_open
[71314.406617] amdair 0000:c1:00.0: amdair_ioctl_alloc_device_memory: Created buffer object, handle 0, mmap offset c000000000000000, size 800000
[71314.406634] amdair 0000:c1:00.0: amdair_mmap: offset c000000000000000
[71314.406640] amdair 0000:c1:00.0: amdair_mmap: Mapping BO with handle 0
[71314.407610] amdair 0000:c1:00.0: Assigning doorbell page 1
[71314.407620] amdair amdair: doorbell offset 0, queue offset 4000000800000000, queue_id 1, db_id 0, dev id 0 DRAM heap CPU VA 75274e200000
[71314.407634] amdair 0000:c1:00.0: amdair_mmap: offset 0
[71314.407653] amdair 0000:c1:00.0: amdair_mmap: offset 4000000800000000
[71314.407667] amdair 0000:c1:00.0: amdair_mmap: offset 8000000800000000

[eddierichter-amd]

Ah great! So, one thing I am confused about is passing device_id=2. It should only be checking for devices that are under control of the AIR driver so you should pass device_id=0. Do you know how you were able to get past the hsa_init failed issue? That is happening because for some reason hsa_init() (https://github.com/Xilinx/mlir-aie/blob/main/runtime_lib/test_lib/test_library.cpp#L181C26-L181C34) is failing when you initialize the device. I don't see how passing a different device_id impacts that, did something else change in the system or application to get that to work?

One thing I would make sure to do in-between failed runs is resetting the device using this script: https://github.com/Xilinx/ROCm-air-platforms/blob/main/platform/vck5000/utils/reset-vck5000.sh. This script reloads the firmware on running on the ARM and the driver to make sure that neither are in a bad state. This script does NOT require a reboot afterwards.

[shibizhao]

Hi,

Thanks for your detailed reply.

I have changed the device_id to 0. But the vector_vector_add and weather_stencil.exe still cannot execute without any output. Just stuck.

I also execute the reset-vck5000.sh script to reset this device without reboot.

The dmesg shows as:

[  677.316403] amdair amdair: amdair_open
[  677.316420] amdair 0000:af:00.0: amdair_ioctl_alloc_device_memory: Created buffer object, handle 0, mmap offset c000000000000000, size 800000
[  677.316428] amdair 0000:af:00.0: amdair_mmap: offset c000000000000000
[  677.316430] amdair 0000:af:00.0: amdair_mmap: Mapping BO with handle 0
[  677.316901] amdair 0000:af:00.0: Assigning doorbell page 1
[  677.316905] amdair amdair: doorbell offset 0, queue offset 4000000800000000, queue_id 1, db_id 0, dev id 0 DRAM heap CPU VA 7fb637d53000
[  677.316911] amdair 0000:af:00.0: amdair_mmap: offset 0
[  677.316921] amdair 0000:af:00.0: amdair_mmap: offset 4000000800000000
[  677.316990] amdair 0000:af:00.0: amdair_mmap: offset 8000000800000000
[  681.890187] amdair amdair: amdair_release
[  733.332596] amdair 0000:af:00.0: removed
[  733.418778] amdair 0000:af:00.0: VCK5000 device found
[  733.418787] amdair 0000:af:00.0: DRAM BAR 0 0x39be00000000 (0x200000000)
[  733.418794] amdair 0000:af:00.0: BRAM BAR 2 0xffffada025d00000 (0x100000)
[  733.418805] amdair 0000:af:00.0: vck5000_init_queues: Initializing queues, queue map ffffffffffffff81 queue size 7000, queue buf size 7000
[  733.418814] amdair 0000:af:00.0: vck5000_init_doorbells: Initializing doorbells, size 7000, doorbell map ffffffffffffff81, num db pages 7
[  733.419229] amdair 0000:af:00.0: Adding AIE 0
[  776.192108] amdair amdair: amdair_open
[  776.192126] amdair 0000:af:00.0: amdair_ioctl_alloc_device_memory: Created buffer object, handle 0, mmap offset c000000000000000, size 800000
[  776.192134] amdair 0000:af:00.0: amdair_mmap: offset c000000000000000
[  776.192136] amdair 0000:af:00.0: amdair_mmap: Mapping BO with handle 0
[  776.192596] amdair 0000:af:00.0: Assigning doorbell page 1
[  776.192600] amdair amdair: doorbell offset 0, queue offset 4000000800000000, queue_id 1, db_id 0, dev id 0 DRAM heap CPU VA 7f9a169af000
[  776.192605] amdair 0000:af:00.0: amdair_mmap: offset 0
[  776.192615] amdair 0000:af:00.0: amdair_mmap: offset 4000000800000000
[  776.192621] amdair 0000:af:00.0: amdair_mmap: offset 8000000800000000
[  788.442889] amdair amdair: amdair_release
[  812.981650] amdair 0000:af:00.0: removed
[  813.053068] amdair 0000:af:00.0: VCK5000 device found
[  813.053077] amdair 0000:af:00.0: DRAM BAR 0 0x39be00000000 (0x200000000)
[  813.053084] amdair 0000:af:00.0: BRAM BAR 2 0xffffada025600000 (0x100000)
[  813.053095] amdair 0000:af:00.0: vck5000_init_queues: Initializing queues, queue map ffffffffffffff81 queue size 7000, queue buf size 7000
[  813.053104] amdair 0000:af:00.0: vck5000_init_doorbells: Initializing doorbells, size 7000, doorbell map ffffffffffffff81, num db pages 7
[  813.053517] amdair 0000:af:00.0: Adding AIE 0
[  834.834176] amdair amdair: amdair_open
[  834.834202] amdair 0000:af:00.0: amdair_ioctl_alloc_device_memory: Created buffer object, handle 0, mmap offset c000000000000000, size 800000
[  834.834211] amdair 0000:af:00.0: amdair_mmap: offset c000000000000000
[  834.834213] amdair 0000:af:00.0: amdair_mmap: Mapping BO with handle 0
[  834.834758] amdair 0000:af:00.0: Assigning doorbell page 1
[  834.834762] amdair amdair: doorbell offset 0, queue offset 4000000800000000, queue_id 1, db_id 0, dev id 0 DRAM heap CPU VA 7fa9cfa09000
[  834.834769] amdair 0000:af:00.0: amdair_mmap: offset 0
[  834.834801] amdair 0000:af:00.0: amdair_mmap: offset 4000000800000000
[  834.834813] amdair 0000:af:00.0: amdair_mmap: offset 8000000800000000

(sandbox) (mlir-aie-dev) root@server:/rshome/bizhao.shi/research/compiler/mlir-aie/build/ROCm-air-platforms/examples/sparta-weather-stencil# ./weather_stencil.exe
Found an AIE HSA agent
[NO ANY OTHER OUTPUT]

This is the output of lspci -vd 10ee: command:

af:00.0 Memory controller: Xilinx Corporation Device b034
        Subsystem: Xilinx Corporation Device 0007
        Flags: fast devsel, IRQ 480, NUMA node 1
        Memory at 39be00000000 (64-bit, prefetchable) [size=8G]
        Memory at ee600000 (64-bit, non-prefetchable) [size=1M]
        Capabilities: [40] Power Management version 3
        Capabilities: [70] Express Endpoint, MSI 00
        Capabilities: [100] Advanced Error Reporting
        Capabilities: [1c0] Secondary PCI Express
        Capabilities: [1f0] Virtual Channel
        Kernel driver in use: amdair

Oh, it is worth to mention that:
when I cold reboot this server, the vck5000 will return to the xilinx qdma shell.

[eddierichter-amd]

Thanks for the detailed response. Think we are getting close. Couple questions:

• After you load the PDI, do you perform a warm reboot? It is possible that the commands are not going over PCIe if the link is not initialized.
• What is the output when you run reset-vck5000.sh? Does it say anything about loading the firmware successfully or unsuccessfully?
• We have a shell running on the ARM, are you able to use minicom or screen to connect to the UART connected to the ARM (on my machine the VCK5000 has /dev/ttyUSB[0-3] and the ARM UART is at /dev/ttyUSB2 and use a baud rate of 115200) and see if it prints anything? Since there is a shell running on the ARM Pressing Enter should return a shell. If you press Enter and nothing comes up that means the firmware was either never loaded or has crashed.