Function testMatmul

experimental/legacy/transformer/test_kernels.cpp:71–107

}

void testMatmul(Context &ctx) {
  static constexpr size_t M = 4;
  static constexpr size_t K = 5;
  static constexpr size_t N = 4;
  auto gen = std::mt19937(31415);
  std::array<float, M * K> input1Arr;
  std::array<float, K * N> input2Arr;
  std::array<float, M * N> outputArr;
  randint(input1Arr, gen, 0, 5);
  range(input2Arr);
  Tensor input1 = createTensor(ctx, {M, K}, kf32, input1Arr.data());
  Tensor input2 = createTensor(ctx, {K, N}, kf32, input2Arr.data());
  Tensor output = createTensor(ctx, {M, N}, kf32, outputArr.data());
  Kernel op = createKernel(
      ctx, MatmulShader(256, kShaderMatMul1, kf32, M, K, N),
      Bindings{input1, input2, output}, {cdiv(M * N, 256), 1, 1});
  std::promise<void> promise;
  std::future<void> future = promise.get_future();
  dispatchKernel(ctx, op, promise);
  wait(ctx, future);
  toCPU(ctx, output, outputArr.data(), sizeof(outputArr));
  LOG(kDefLog, kInfo, "%s", show<float, M, K>(input1Arr, "A").c_str());
  LOG(kDefLog, kInfo, "%s", show<float, K, N>(input2Arr, "B").c_str());
  LOG(kDefLog, kInfo, "%s", show<float, M, N>(outputArr, "C").c_str());

  std::array<float, M * N> refOutputArr;
  std::array<float, K * N> input2ArrT;
  transpose(input2Arr.data(), input2ArrT.data(), K, N);
  LOG(kDefLog, kInfo, "%s", show<float, N, K>(input2ArrT, "B'").c_str());
  ref::matmul_forward_cpu(refOutputArr.data(), input1Arr.data(),
                          input2ArrT.data(), nullptr, 1, M, K, N);
  LOG(kDefLog, kInfo, show<float, M, N>(refOutputArr, "C (reference)").c_str());

  LOG(kDefLog, kInfo, "Done with Matmul Test");
  bool passed = isclose(outputArr.data(), refOutputArr.data(), N);
  assert(passed);
}

void testTensorPool(Context &ctx) {
  LOG(kDefLog, kInfo, "Starting Tensor Pool Test");