Function createTranspose2

examples/transpose/run.cpp:81–103

)";

inline KernelCode createTranspose2(const char *shaderTemplate,
				   const size_t M, const size_t N,
				   const size_t BM, const size_t BN,
                                   const size_t TM, const size_t TN,
                                   const Shape &workgroupSize = {256, 1, 1},
                                   NumType precision = kf32) {
  assert(BM % TM == 0);
  assert(BN % TN == 0);
  assert(M % BM == 0);
  assert(N % BN == 0);
  std::string codeString(shaderTemplate);
  replaceAll(codeString, {{"{{workgroupSize}}", toString(workgroupSize)},
                          {"{{precision}}", toString(precision)},
                          {"{{M}}", toString(M)},
                          {"{{N}}", toString(N)},
                          {"{{BM}}", toString(BM)},
                          {"{{BN}}", toString(BN)},
                          {"{{TM}}", toString(TM)},
                          {"{{TN}}", toString(TN)}
                          });
  std::string unrolledCode = codeString ;// loopUnrolling(codeString);
  return {unrolledCode, workgroupSize};
}

void initData(size_t M, size_t N, std::unique_ptr<float[]> &inputPtr) {
  std::mt19937 gen(314159);