XVDPU-TRD C64B5 timeout on VCK190-ES

[ksstms]

I've modified the XVDPU-TRD to target the VCK190-ES board, and removed the HDMI and MIPI stuff from the design.
Everything works fine, except for the C64-B5 DPU configuration.

I get setup violations inside the DPU. Unfortunately I don't know how I could provide more info about it, as the path endpoints are <hidden>. On the schematic I can see that most of the paths are around URAM instances, and some of them are coming from a NOC AXI port. The clock is clkout1_primitive_1 which is the 150 MHz clock connected to s_axi_aclk of the DPU.

Every other configuration works with the original 150 MHz clock. PG389 page 30 recommends using 100 MHz for s_axi_aclk, so I modified that in the Makefile:

@@ -107,7 +107,7 @@ $(BUILD_DIR)/binary_container_1.xclbin: $(BINARY_CONTAINER_1_OBJS)
        @$(VXX) $(VXXFLAGS) -l --config scripts/system.cfg --config scripts/xvdpu_aie_noc.cfg \
          --clock.freqHz $(PL_FREQ):DPUCVDX8G_1.m_axi_aclk \
          --clock.freqHz $(PL_FREQ):ai_engine_0.aclk0 \
-         --clock.freqHz 150000000:DPUCVDX8G_1.s_axi_aclk \
+         --clock.freqHz 100000000:DPUCVDX8G_1.s_axi_aclk \
          -o "$@" $(BINARY_CONTAINER_1_OBJS) 

Now the timing is okay, but when I try to benchmark it, I get the following error:

root@xilinx-vck190-2021_2:~# xdputil benchmark resnet50-64-5.xmodel -i -1 2
XAIEFAL: INFO: Resource group Avail is created.
XAIEFAL: INFO: Resource group Static is created.
XAIEFAL: INFO: Resource group Generic is created.
WARNING: Logging before InitGoogleLogging() is written to STDERR
I0208 04:14:22.811815  1070 test_dpu_runner_mt.cpp:475] shuffle results for batch...
I0208 04:14:22.814556  1070 performance_test.hpp:73] 0% ...
I0208 04:14:28.814703  1070 performance_test.hpp:76] 10% ...
W0208 04:14:33.576429  1078 xrt_cu.cpp:188] cu timeout! device_core_idx 0  handle=0xffffa0010a10 ENV_PARAM(XLNX_DPU_TIMEOUT) 10000 state 1 ERT_CMD_STATE_COMPLETED 4 ms 10010  bo=1 is_done 0 
I0208 04:14:33.576498  1078 xrt_cu.cpp:99] Total: 10010162us	ToDriver: 18446744073705911us	ToCU: 1254us	Complete: 1808us	Done: 10010739us
F0208 04:14:33.576524  1078 dpu_control_xrt_xv_dpu.cpp:193] dpu timeout! core_idx = 0
 AP 1  LSTART 309  LEND 309  CSTART 217  CEND 215  SSTART 0  SEND 0  MSTART 214  MEND 214  CYCLE_L 1000905996  CYCLE_H 0 
*** Check failure stack trace: ***
/usr/bin/xdputil: line 20:  1068 Aborted                 /usr/bin/python3 -m xdputil $*

This is the same error that I got in issue #576. Is the ES AIE workaround script wrong for this configuration?

[ksstms]

Since this configuration uses more AIE cores than any other, I got suspicious of the ES AIE workaround script.
I checked it, and found that it only sets registers in cores from 0_0 to 39_8. The 4x_x cores are not affected.

In the script below I tried to modify the outer for loop to run to 49 instead of 39, and now the benchmark runs fine.

Vitis-AI/dsa/XVDPU-TRD/README.md

Lines 317 to 325 in fc74d40

	for i in {0..39}
	do
	for j in {1..8}
	do
	a=0x20000000000
	b=0x31000
	devmem $[a+b+(i<<23)+(j<<18)] 32 0
	done
	done

Please let me know if this is actually a good fix.

[qianglin-xlnx]

Hi @ksstms
Thank you for your feedback. It's a good fix. We will update the script in README.md

[ksstms]

Great! Thanks!

[ksstms]

dsa/XVDPU-TRD/README.md fixed by 6b96cc3