poyenc

SRC = main.cpp
ARCH = gfx942
INC_DIR = include/ck_tile

all:
	clang++ -fsyntax-only -I$(INC_DIR) -x hip --offload-arch=$(ARCH) $(SRC)

poyenc

#include <hip/hiprtc.h>
#include <hip/hip_runtime.h>

#include <cstdlib>
#include <iostream>
#include <string>
#include <vector>

static constexpr const char gpu_program[]{
R"(

poyenc

# run packed docker image
docker run              \
-it                     \
--privileged            \
--group-add sudo        \
poyenc/9110-mod:latest

# configure CK
cd ~/composable_kernel
mkdir build && cd build

poyenc

#include <array>
#include <cassert>
#include <chrono>
#include <cstdio>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>

poyenc

// CodeEmitVisitor.h
class CodeEmitVisitor : public CustomVisitor<CodeEmitVisitor>, private NvDlaConstants
{
public:
   void visit(const Conv& pConv) override;
   void visit(Conv& pConv) override;
   
   // 1. add visit method for const Add IR
   void visit(const Add& pOp) override;   
   // 2. add visit method for non-const Add IR, force it calls the previous version

poyenc

#pragma once                                                                                                                                         [33/1792]

#include <algorithm>
#include <cstddef>
#include <deque>
#include <functional>
#include <limits>
#include <random>
#include <type_traits>
#include <utility>

poyenc

#pragma once                                                                                                                                         [27/1718]

#include <iterator>
#include <memory>
#include <type_traits>
#include <utility>

namespace poyenc {

template <typename Iterator, typename Generator>

poyenc

#pragma once

#include <stdexcept>
#include <type_traits>
#include <utility>

namespace poyenc {

template <typename IntType = int>
class sequence_generator {

poyenc

#pragma once

#include <iterator>
#include <type_traits>
#include <utility>

template <typename RandomAccessRange>
class index_iterator {
public:
  using target_type         = RandomAccessRange;

poyenc

#include <cassert>
#include <cstddef>
#include <iostream>
#include <type_traits>
#include <utility>

namespace poyenc {
namespace mpl {

template <typename...>