Merlin-UCLA

Use Merlin-UCLA with Docker

Download

docker pull ghcr.io/ucla-vast/merlin-ucla:latest

Configuration

Please update the file run_docker.sh with you own paths.

Run in interactive mode

sh run_docker.sh

# Please change these paths
xilinx_path=/opt/xilinx
tools_path=/opt/tools
XILINX_VITIS=/opt/tools/xilinx/Vitis/2021.1
XILINX_XRT=/opt/xilinx/xrt
XILINX_VIVADO=/opt/tools/xilinx/Vivado/2021.1
LM_LICENSE_FILE=

########################

CURRENT_PATH=$(cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd)
imagename="ghcr.io/ucla-vast/merlin-ucla:latest"

docker run -v /home:/home -v $tools_path:$tools_path -v $xilinx_path:$xilinx_path -e XILINX_VITIS=$XILINX_VITIS -e XILINX_XRT=$XILINX_XRT -e XILINX_VIVADO=$XILINX_VIVADO -e LM_LICENSE_FILE=$LM_LICENSE_FILE -w="$CURRENT_PATH" -it "$imagename"

Install Merlin-UCLA

Prequest dependency environment:

python >= 3.6.8
cmake >= 3.19.0
boost == 1.67.0
clang == 6.0.0
gcc == 4.9.4
llvm == 6.0.0

How to build:

In merlin_setting.sh, change MERLIN_COMPILER_HOME to your absolute path
Download gcc4.9.4 to $gcc_path specified in merlin_settings.sh and compile, add built library path to LD_LIBRARY_PATH in merlin_setting.sh
Download llvm6.0.0, clang 6.0.0, boost1.67.0 to the path specified in merlin_settings.sh, and compile all the packages with gcc4.9.4 built in step 2
source merlin_setting.sh
cd trunk/build;
cmake3 -DCMAKE_BUILD_TYPE=Release ..;
make -j;

Run Merlin-UCLA

Please first source all the necessary paths e.g., source /opt/tools/xilinx/Vitis_HLS/2021.1/settings64.sh.

Run

merlincc

To print the help:

merlincc -h

Compilation Options

To select the platform please use the option:

--platform=<the platform>
#for example
--platform=vitis::/opt/xilinx/platforms/xilinx_u200_xdma_201830_2/xilinx_u200_xdma_201830_2.xpfm

To change frequency please add the option:

--kernel_frequency <frequency in MHz>
# for example
--kernel_frequency 250

To automatically do tree reduction in logarithmic time please add the option:

-funsafe-math-optimizations

To change the burst single size threshold please add the option:

--attribute burst_single_size_threshold=<size>
# for example
--attribute burst_single_size_threshold=36700160

To change the burst total size threshold please add the option:

--attribute burst_total_size_threshold=<size>
# for example
--attribute burst_total_size_threshold=36700160

You can include the path of the include folder with the option -I. For example we use the options:

CFLAGS="-I $XILINX_HLS/include" merlincc --attribute burst_total_size_threshold=36700160 --attribute burst_single_size_threshold=36700160 --kernel_frequency 250  -funsafe-math-optimizations --platform=vitis::/opt/xilinx/platforms/xilinx_u200_xdma_201830_2/xilinx_u200_xdma_201830_2.xpfm  -I $XILINX_HLS/lnx64/tools/gcc/lib/gcc/x86_64-unknown-linux-gnu/4.6.3/include/ -I $XILINX_HLS/include/ -I /opt/merlin/sources/merlin-compiler/trunk/source-opt/include/apint_include/ -c -o mykernel_merlincc_polyopt --report=estimate

Pragmas

All pragmas are applied above the loop contrary to Xilinx Vitis.

Mandatory Pragma

Please add #pragma ACCEL kernel above the main function. For example:

#pragma ACCEL kernel
void kernel_bicg(int m,int n,float A[2100][1900],float s[1900],float q[2100],float p[1900],float r[2100])
{
  int i;
  int j;    
    for (i = 0; i < 1900; i++) {
      s[i] = 0;
    }
    for (i = 0; i < 2100; i++) {
      q[i] = 0.0;
      for (j = 0; j < 1900; j++) {
        s[j] = s[j] + r[i] * A[i][j];
        q[i] = q[i] + A[i][j] * p[j];
      }
    }
}

Hardware directives

Parallel / Unroll

#pragma ACCEL parallel factor=<uf>

If factor is not specify the loop is fully unrolled.

Pipeline

#pragma ACCEL pipeline flatten

The loop bellow the pragma will be pipeline and the innermost loop will be fully unrolled.

Note: #pragma ACCEL pipeline flatten can also be use as #pragma ACCEL pipeline if all the innermost loop are fully unrolled.

Tile 1D / Strip mining

#pragma ACCEL tile factor=<tile size>

The loop bellow the pragma will be strip mined and the innermost loop create will have a trip count equal to tile size.

Memory communication

#pragma ACCEL cache variable=<name array>

The transfer from off-chip to on-chip will be done at the position of the pragma cache.

Double buffer

#pragma ACCEL pipeline

The communication of the array will be done at the position of the pragma and transfer with double buffer.

UCLA-VAST/Merlin-UCLA