Run the Application Using DDR - 2023.1 English - XD099

Vitis Tutorials: Hardware Acceleration (XD099)

Document ID
XD099
Release Date
2023-08-02
Version
2023.1 English

Run the hardware application using DDR with a size of 600 MB, sequential address pattern, and enqueue the kernel one time. The host will migrate 600 MB to DDR0 (buffer_in1) and DDR1(buffer_in2) respectively. The kernel will perform the compute and store the results in DDR2, buffer_output.

Here is the makefile command to run ().

#make ddr_addSeq_build  - executed already in first module.
make ddr_addSeq

The above run command essentially expands to the following.

make run TARGET=hw memtype=DDR dsize=600 addrndm=0 krnl_loop=1 buildxclbin=0
  • memtype: Sets memory as DDR or HBM.

  • dsize: The amount of data migrated by the host to memory banks and accessed by the kernel ports, in1 and in2.

  • kernel_loop: Sets the number of time the kernel loop repeats.

  • buildxclbin=0: Will not generate the new xclbin.

  • txSize: Set to 64 by default. It is the size of transactions issued by kernel port while accessing memory.

The preceding make command will create the build directory ../build/DDR_Banks_d512_txSize64 reflecting the memory type, data and transaction size.

The above commands to run the application on hardware show the following results:

*** Running hw mode ***  Use Command Line to run application!
cd ./../build/DDR_Banks_d512_txSize64 &&  ./host vadd_hw.xclbin 600 0 1 64;

 Total Data of 600.000 Mbytes to be written to global memory from host

 Kernel is invoked 1 time and repeats itself 1 times

Found Platform
Platform Name: Xilinx
DEVICE xilinx_u250_gen3x16_xdma_4_1_202210_1
INFO: Reading vadd_hw.xclbin
Loading: 'vadd_hw.xclbin'
- host loop iteration #0 of 1 total iterations
kernel_time_in_sec = 0.0416315
Duration using events profiling: 41473086 ns
 match_count = 157286400 mismatch_count = 0 total_data_size = 157286400
Throughput Achieved = 15.17 GB/s
TEST PASSED

The host is migrating 600 MB of data to both DDR0 and DDR1. The kernel accesses this data using in1 and in2 ports from DDR0 and DDR1, respectively. The vector addition is performed by the kernel, and the results are written to DDR2. These results from DDR2 are migrated back to the host. The next section goes over the steps required to migrate this DDR-based application to HBM.

TIP: TARGET=hw_emu can also be used for running hardware emulation, but this will take significant time to run the application for a 600 MB size buffer.