Function testResidual

experimental/legacy/transformer/test_kernels.cpp:15–45

using namespace gpu;

void testResidual(Context &ctx) {
  constexpr size_t N = 200000;
  constexpr size_t workgroupSize = 256;
  std::array<float, N> input1Arr;
  std::array<float, N> input2Arr;
  range(input1Arr);
  range(input2Arr);
  std::array<float, N> outputArr;
  Tensor input1 = createTensor(ctx, {N}, kf32, input1Arr.data());
  Tensor input2 = createTensor(ctx, {N}, kf32, input2Arr.data());
  Tensor output = createTensor(ctx, {N}, kf32, outputArr.data());
  std::promise<void> promise;
  std::future<void> future = promise.get_future();
  KernelCode shaderCode = {kShaderResidual, workgroupSize, kf32};
  LOG(kDefLog, kInfo, "Shader Code :\n%s", shaderCode.data.c_str());
  Kernel op =
      createKernel(ctx, {kShaderResidual, workgroupSize, kf32},
                   Bindings{input1, input2, output}, {cdiv(N, workgroupSize), 1, 1});
  dispatchKernel(ctx, op, promise);
  wait(ctx, future);
  toCPU(ctx, output, outputArr.data(), sizeof(outputArr));
  LOG(kDefLog, kInfo, "%s",
      show<float, N, 1>(outputArr, "Residual Output").c_str());
  std::array<float, N> outputRef;
  ref::residual_forward_cpu(outputRef.data(), input1Arr.data(),
                            input2Arr.data(), N);
  LOG(kDefLog, kInfo, "%s",
      show<float, N, 1>(outputRef, "Residual Reference Output").c_str());
  assert(isclose(outputArr.data(), outputRef.data(), N));
  LOG(kDefLog, kInfo, "Done with Residual Test");
}

void testHadamard(Context &ctx) {
  constexpr size_t N = 200000;