eager_py_layer.cc 18.2 KB
Newer Older
W
wanghuancoder 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
// disable numpy compile error
#include <Python.h>

#include <set>
#include <string>
#include <vector>

#pragma GCC diagnostic ignored "-Wattributes"
#include "pybind11/pytypes.h"

#include "paddle/fluid/eager/accumulation/accumulation_node.h"
#include "paddle/fluid/eager/api/all.h"
#include "paddle/fluid/eager/autograd_meta.h"
#include "paddle/fluid/eager/pylayer/py_layer_node.h"
#include "paddle/fluid/eager/utils.h"
#include "paddle/fluid/framework/convert_utils.h"
#include "paddle/fluid/memory/allocation/allocator.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/pybind/eager.h"
#include "paddle/fluid/pybind/eager_utils.h"
#include "paddle/fluid/pybind/exception.h"
#include "paddle/phi/common/data_type.h"
#include "paddle/phi/core/compat/convert_utils.h"
#include "paddle/phi/core/dense_tensor.h"
#include "pybind11/detail/internals.h"
37 38
#pragma GCC diagnostic ignored "-Wwrite-strings"
#pragma GCC diagnostic ignored "-Wmissing-field-initializers"
W
wanghuancoder 已提交
39 40 41 42 43 44 45 46 47

namespace paddle {
namespace pybind {

namespace py = ::pybind11;

PyTypeObject* p_pylayer_type;
extern PyTypeObject* p_tensor_type;

48
std::set<paddle::experimental::Tensor*> GetTensorsFromPyObject(PyObject* obj) {
W
wanghuancoder 已提交
49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
  std::set<paddle::experimental::Tensor*> result;
  if (obj == nullptr) {
    return result;
  }
  if (IsEagerTensor(obj)) {
    result.insert(&reinterpret_cast<TensorObject*>(obj)->tensor);  // NOLINT
  } else if (PyList_Check(obj)) {
    Py_ssize_t len = PyList_Size(obj);
    for (Py_ssize_t i = 0; i < len; i++) {
      if (IsEagerTensor(PyList_GetItem(obj, i))) {
        result.insert(
            &reinterpret_cast<TensorObject*>(PyList_GetItem(obj, i))  // NOLINT
                 ->tensor);
      }
    }
  } else if (PyTuple_Check(obj)) {
    Py_ssize_t len = PyTuple_Size(obj);
    for (Py_ssize_t i = 0; i < len; i++) {
      if (IsEagerTensor(PyTuple_GetItem(obj, i))) {
        result.insert(
            &reinterpret_cast<TensorObject*>(PyTuple_GetItem(obj, i))  // NOLINT
                 ->tensor);
      }
    }
  }
  return result;
}

PyObject* PyLayerNew(PyTypeObject* type, PyObject* args, PyObject* kwargs) {
  PyObject* obj = type->tp_alloc(type, 0);
  if (obj) {
    auto v = reinterpret_cast<PyLayerObject*>(obj);
    v->materialize_grads = true;
    new (&v->grad_node) std::weak_ptr<egr::GradNodePyLayer>();
    new (&v->forward_input_tensor_is_duplicable) std::vector<bool>();
    new (&v->forward_output_tensor_is_duplicable) std::vector<bool>();
  }
  return obj;
}

static void PyLayerDealloc(PyLayerObject* self) {
  if (self->container) {
    Py_DECREF(self->container);
  }
  if (self->non_differentiable) {
    Py_DECREF(self->non_differentiable);
  }
  if (self->dirty_tensors) {
    Py_DECREF(self->dirty_tensors);
  }
  self->grad_node.~weak_ptr<egr::GradNodePyLayer>();
  self->forward_input_tensor_is_duplicable.~vector();
  self->forward_output_tensor_is_duplicable.~vector();
  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
}

PyObject* pylayer_method_name(PyObject* self, PyObject* noargs) {
  EAGER_TRY
  return ToPyObject(
      reinterpret_cast<PyLayerObject*>(self)->grad_node.lock()->name());
  EAGER_CATCH_AND_THROW_RETURN_NULL
}

PyObject* pylayer_method_apply(PyObject* cls, PyObject* args,
                               PyObject* kwargs) {
  EAGER_TRY
  VLOG(6) << "Begin run PyLayer apply...";
  PyObject* backward_function =
      PyObject_GetAttrString(cls, "_backward_function");
  if (!backward_function) {
    PADDLE_THROW(paddle::platform::errors::InvalidArgument(
        "Get _backward_function faild."));
  }
  PyLayerObject* ctx = reinterpret_cast<PyLayerObject*>(
      PyObject_CallFunctionObjArgs(backward_function, nullptr));
  if (!ctx) {
    return nullptr;
  }
  VLOG(6) << "PyLayer construct PyLayerContext finish...";

  bool require_any_grad = false;

  size_t inputs_size = 0;
  PyObject* forward_args = nullptr;
  PyObject* kwargs_value_list = nullptr;
  if (kwargs) {
    inputs_size = PyDict_Size(kwargs);
    kwargs_value_list = PyDict_Values(kwargs);
    forward_args = PyTuple_New(1);
  } else {
    inputs_size = PyTuple_GET_SIZE(args);
    forward_args = PyTuple_New(inputs_size + 1);
  }
  Py_INCREF(ctx);
  PyTuple_SET_ITEM(forward_args, 0, reinterpret_cast<PyObject*>(ctx));

  std::vector<std::vector<egr::AutogradMeta*>> inputs_autograd_meta;
  inputs_autograd_meta.reserve(inputs_size);
  std::vector<std::vector<paddle::experimental::Tensor*>> inputs_tensor;
  inputs_tensor.reserve(inputs_size);
  ctx->forward_input_tensor_is_duplicable.clear();
  ctx->forward_input_tensor_is_duplicable.reserve(inputs_size);
  for (size_t i = 0; i < inputs_size; i++) {
    PyObject* obj = nullptr;
    if (kwargs) {
      obj = PyList_GetItem(kwargs_value_list, i);
    } else {
      obj = PyTuple_GET_ITEM(args, i);
    }
    if (IsEagerTensor(obj)) {
      auto autograd_meta = egr::EagerUtils::nullable_autograd_meta(
          reinterpret_cast<TensorObject*>(obj)->tensor);
      inputs_autograd_meta.push_back({autograd_meta});
      inputs_tensor.push_back(
          {&(reinterpret_cast<TensorObject*>(obj)->tensor)});  // NOLINT
      bool stop_gradient =
          autograd_meta == nullptr ? true : autograd_meta->StopGradient();
      if (!stop_gradient) {
        require_any_grad = true;
      }
      ctx->forward_input_tensor_is_duplicable.push_back(false);
    } else if (PyList_Check(obj)) {
      std::vector<paddle::experimental::Tensor*> tensors;
      Py_ssize_t len = PyList_Size(obj);
      for (Py_ssize_t i = 0; i < len; i++) {
        if (IsEagerTensor(PyList_GetItem(obj, i))) {
          tensors.push_back(&(
              reinterpret_cast<TensorObject*>(PyList_GetItem(obj, i))->tensor));
        }
      }
      if (!tensors.empty()) {
        auto autograd_meta = egr::EagerUtils::nullable_autograd_meta(tensors);
        for (auto iter : autograd_meta) {
          bool stop_gradient = iter == nullptr ? true : iter->StopGradient();
          if (!stop_gradient) {
            require_any_grad = true;
          }
        }
        inputs_autograd_meta.push_back(autograd_meta);
        inputs_tensor.push_back(tensors);
        ctx->forward_input_tensor_is_duplicable.push_back(true);
      }
    } else if (PyTuple_Check(obj)) {
      std::vector<paddle::experimental::Tensor*> tensors;
      Py_ssize_t len = PyTuple_Size(obj);
      for (Py_ssize_t i = 0; i < len; i++) {
        if (IsEagerTensor(PyTuple_GetItem(obj, i))) {
          tensors.push_back(
              &(reinterpret_cast<TensorObject*>(PyTuple_GetItem(obj, i))
                    ->tensor));
        }
      }
      if (!tensors.empty()) {
        auto autograd_meta = egr::EagerUtils::nullable_autograd_meta(tensors);
        for (auto iter : autograd_meta) {
          bool stop_gradient = iter == nullptr ? true : iter->StopGradient();
          if (!stop_gradient) {
            require_any_grad = true;
          }
        }
        inputs_autograd_meta.push_back(autograd_meta);
        inputs_tensor.push_back(tensors);
        ctx->forward_input_tensor_is_duplicable.push_back(true);
      }
    }

    if (!kwargs) {
      Py_INCREF(obj);
      PyTuple_SET_ITEM(forward_args, i + 1, obj);
    }
  }

  VLOG(6)
      << "PyLayer forward args is ready, begin call user's forward function...";
  // call forward
  auto forward_fn = PyObject_GetAttrString(cls, "forward");
  if (!forward_fn) {
    PADDLE_THROW(paddle::platform::errors::InvalidArgument(
        "Get forward function faild."));
  }
  bool trace_backward = egr::Controller::Instance().HasGrad();
  egr::Controller::Instance().SetHasGrad(false);
  auto outputs = PyObject_Call(forward_fn, forward_args, kwargs);
  egr::Controller::Instance().SetHasGrad(trace_backward);
  if (!outputs) {
234 235 236 237
    Py_XDECREF(forward_args);
    Py_XDECREF(kwargs_value_list);
    Py_XDECREF(backward_function);
    Py_XDECREF(forward_fn);
W
wanghuancoder 已提交
238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
    return nullptr;
  }

  PyObject* outputs_tuple = nullptr;
  if (PyTuple_Check(outputs)) {
    outputs_tuple = outputs;
  } else {
    outputs_tuple = PyTuple_New(1);
    Py_INCREF(outputs);
    PyTuple_SET_ITEM(outputs_tuple, 0, outputs);
  }

  auto outputs_size = PyTuple_GET_SIZE(outputs_tuple);
  std::vector<std::vector<paddle::experimental::Tensor*>> outputs_tensor;
  outputs_tensor.reserve(outputs_size);
  std::vector<std::vector<egr::AutogradMeta*>> outputs_autograd_meta;
  outputs_autograd_meta.reserve(outputs_size);
  ctx->forward_output_tensor_is_duplicable.clear();
  ctx->forward_output_tensor_is_duplicable.reserve(outputs_size);
  for (Py_ssize_t i = 0; i < outputs_size; i++) {
    PyObject* obj = PyTuple_GET_ITEM(outputs_tuple, i);
    if (IsEagerTensor(obj)) {
      outputs_tensor.push_back(
          {&(reinterpret_cast<TensorObject*>(obj)->tensor)});
      outputs_autograd_meta.push_back({egr::EagerUtils::autograd_meta(
          &(reinterpret_cast<TensorObject*>(obj)->tensor))});
      ctx->forward_output_tensor_is_duplicable.push_back(false);
    } else if (PyList_Check(obj)) {
      std::vector<paddle::experimental::Tensor*> tensors;
      Py_ssize_t len = PyList_Size(obj);
      for (Py_ssize_t i = 0; i < len; i++) {
        if (IsEagerTensor(PyList_GetItem(obj, i))) {
          tensors.push_back(&(
              reinterpret_cast<TensorObject*>(PyList_GetItem(obj, i))->tensor));
        }
      }
      if (!tensors.empty()) {
        outputs_tensor.push_back(tensors);
        outputs_autograd_meta.push_back(
            egr::EagerUtils::autograd_meta(&tensors));
        ctx->forward_output_tensor_is_duplicable.push_back(true);
      }
    } else if (PyTuple_Check(obj)) {
      std::vector<paddle::experimental::Tensor*> tensors;
      Py_ssize_t len = PyTuple_Size(obj);
      for (Py_ssize_t i = 0; i < len; i++) {
        if (IsEagerTensor(PyTuple_GetItem(obj, i))) {
          tensors.push_back(
              &(reinterpret_cast<TensorObject*>(PyTuple_GetItem(obj, i))
                    ->tensor));
        }
      }
      if (!tensors.empty()) {
        outputs_tensor.push_back(tensors);
        outputs_autograd_meta.push_back(
            egr::EagerUtils::autograd_meta(&tensors));
        ctx->forward_output_tensor_is_duplicable.push_back(true);
      }
    }
  }

  if (outputs_tensor.size() == 0) {
    PADDLE_THROW(platform::errors::InvalidArgument(
        "At least one output of `PyLayer.forward` is a `Tensor`."));
  }
  VLOG(6) << "PyLayer forward function finish...";

  if (require_any_grad && trace_backward) {
306
    auto non_differentiable = GetTensorsFromPyObject(ctx->non_differentiable);
W
wanghuancoder 已提交
307 308 309 310 311 312 313 314 315 316 317
    for (size_t i = 0; i < outputs_autograd_meta.size(); i++) {
      for (size_t j = 0; j < outputs_autograd_meta[i].size(); j++) {
        if (non_differentiable.find(outputs_tensor[i][j]) !=
            non_differentiable.end()) {
          outputs_autograd_meta[i][j]->SetStopGradient(true);
        } else {
          outputs_autograd_meta[i][j]->WeakSetStopGradient(false);
        }
      }
    }

318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333
    // add inplace strategy, inplaced tensor is ctx->dirty_tensors
    auto dirty_tensors = GetTensorsFromPyObject(ctx->dirty_tensors);
    for (auto it = dirty_tensors.begin(); it != dirty_tensors.end(); ++it) {
      auto dirty_tensor = *it;
      auto dirty_tensor_autograd_meta =
          egr::EagerUtils::autograd_meta(dirty_tensor);
      PADDLE_ENFORCE_EQ(!dirty_tensor_autograd_meta->StopGradient() &&
                            egr::egr_utils_api::IsLeafTensor(*dirty_tensor),
                        false, paddle::platform::errors::InvalidArgument(
                                   "Leaf Var (%s) that doesn't stop gradient "
                                   "can't use inplace strategy.",
                                   dirty_tensor->name()));
      dirty_tensor->bump_inplace_version();
      VLOG(3) << "Tensor(" << dirty_tensor->name()
              << ") uses Inplace Strategy.";
    }
W
wanghuancoder 已提交
334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371

    auto grad_node = std::make_shared<egr::GradNodePyLayer>(
        reinterpret_cast<PyObject*>(ctx), outputs_autograd_meta.size(),
        inputs_autograd_meta.size());
    ctx->grad_node = grad_node;

    if (ctx->materialize_grads) {
      grad_node->SaveForwardOutputsMeta(outputs_tensor);
    }

    for (size_t i = 0; i < inputs_autograd_meta.size(); i++) {
      if (ctx->forward_input_tensor_is_duplicable[i]) {
        for (auto t : inputs_tensor[i]) {
          grad_node->SetGradOutMeta(*t, i);
        }
      } else {
        grad_node->SetGradOutMeta(*inputs_tensor[i][0], i);
      }
    }

    for (size_t i = 0; i < outputs_autograd_meta.size(); i++) {
      if (ctx->forward_output_tensor_is_duplicable[i]) {
        egr::EagerUtils::SetOutRankWithSlot(&outputs_autograd_meta[i], i);
        egr::EagerUtils::SetHistory(&outputs_autograd_meta[i], grad_node);
        for (auto t : outputs_tensor[i]) {
          grad_node->SetGradInMeta(*t, i);
        }
        egr::EagerUtils::CheckAndRetainGrad(outputs_tensor[i]);
      } else {
        egr::EagerUtils::SetOutRankWithSlot(outputs_autograd_meta[i][0], i);
        egr::EagerUtils::SetHistory(outputs_autograd_meta[i][0], grad_node);
        grad_node->SetGradInMeta(*outputs_tensor[i][0], i);
        egr::EagerUtils::CheckAndRetainGrad(*outputs_tensor[i][0]);
      }
    }
    VLOG(6) << "PyLayer construct backward node finish...";
  }

372 373 374 375 376 377 378 379
  if (!PyTuple_Check(outputs)) {
    Py_XDECREF(outputs_tuple);
  }
  Py_XDECREF(forward_args);
  Py_XDECREF(kwargs_value_list);
  Py_XDECREF(backward_function);
  Py_XDECREF(forward_fn);

W
wanghuancoder 已提交
380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407
  return outputs;
  EAGER_CATCH_AND_THROW_RETURN_NULL
}

PyObject* pylayer_method_register_hook(PyObject* _self, PyObject* hook) {
  EAGER_TRY
  return nullptr;
  EAGER_CATCH_AND_THROW_RETURN_NULL
}

PyObject* tensor_properties_get_container(PyLayerObject* self, void* closure) {
  EAGER_TRY
  if (self->container == nullptr) {
    Py_INCREF(Py_None);
    return Py_None;
  }
  Py_INCREF(self->container);
  return self->container;
  EAGER_CATCH_AND_THROW_RETURN_NULL
}

int tensor_properties_set_container(PyLayerObject* self, PyObject* value,
                                    void* closure) {
  EAGER_TRY
  Py_XINCREF(value);
  Py_XDECREF(self->container);
  self->container = value;
  return 0;
0
0x45f 已提交
408
  EAGER_CATCH_AND_THROW_RETURN_NEG
W
wanghuancoder 已提交
409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429
}

PyObject* tensor_properties_get_non_differentiable(PyLayerObject* self,
                                                   void* closure) {
  EAGER_TRY
  if (self->non_differentiable == nullptr) {
    Py_INCREF(Py_None);
    return Py_None;
  }
  Py_INCREF(self->non_differentiable);
  return self->non_differentiable;
  EAGER_CATCH_AND_THROW_RETURN_NULL
}

int tensor_properties_set_non_differentiable(PyLayerObject* self,
                                             PyObject* value, void* closure) {
  EAGER_TRY
  Py_XINCREF(value);
  Py_XDECREF(self->non_differentiable);
  self->non_differentiable = value;
  return 0;
0
0x45f 已提交
430
  EAGER_CATCH_AND_THROW_RETURN_NEG
W
wanghuancoder 已提交
431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451
}

PyObject* tensor_properties_get_dirty_tensors(PyLayerObject* self,
                                              void* closure) {
  EAGER_TRY
  if (self->dirty_tensors == nullptr) {
    Py_INCREF(Py_None);
    return Py_None;
  }
  Py_INCREF(self->dirty_tensors);
  return self->dirty_tensors;
  EAGER_CATCH_AND_THROW_RETURN_NULL
}

int tensor_properties_set_dirty_tensors(PyLayerObject* self, PyObject* value,
                                        void* closure) {
  EAGER_TRY
  Py_XINCREF(value);
  Py_XDECREF(self->dirty_tensors);
  self->dirty_tensors = value;
  return 0;
0
0x45f 已提交
452
  EAGER_CATCH_AND_THROW_RETURN_NEG
W
wanghuancoder 已提交
453 454 455 456 457 458 459
}

int tensor_properties_set_materialize_grads(PyLayerObject* self,
                                            PyObject* value, void* closure) {
  EAGER_TRY
  self->materialize_grads = CastPyArg2AttrBoolean(value, 0);
  return 0;
0
0x45f 已提交
460
  EAGER_CATCH_AND_THROW_RETURN_NEG
W
wanghuancoder 已提交
461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493
}

PyMethodDef pylayer_methods[] = {
    {"name", (PyCFunction)(void (*)(void))pylayer_method_name, METH_NOARGS,
     NULL},
    {"apply", (PyCFunction)(void (*)(void))pylayer_method_apply,
     METH_CLASS | METH_VARARGS | METH_KEYWORDS, NULL},
    {"register_hook", (PyCFunction)(void (*)(void))pylayer_method_register_hook,
     METH_O, NULL},
    {NULL, NULL, 0, NULL}};

struct PyGetSetDef pylayer_properties[]{
    {"container", (getter)tensor_properties_get_container,
     (setter)tensor_properties_set_container, nullptr, nullptr},
    {"non_differentiable", (getter)tensor_properties_get_non_differentiable,
     (setter)tensor_properties_set_non_differentiable, nullptr, nullptr},
    {"dirty_tensors", (getter)tensor_properties_get_dirty_tensors,
     (setter)tensor_properties_set_dirty_tensors, nullptr, nullptr},
    {"materialize_grads", nullptr,
     (setter)tensor_properties_set_materialize_grads, nullptr, nullptr},
    {nullptr, nullptr, nullptr, nullptr, nullptr}};

void BindEagerPyLayer(PyObject* module) {
  auto heap_type = reinterpret_cast<PyHeapTypeObject*>(
      PyType_Type.tp_alloc(&PyType_Type, 0));
  heap_type->ht_name = ToPyObject("PyLayer");
  heap_type->ht_qualname = ToPyObject("PyLayer");
  auto type = &heap_type->ht_type;
  type->tp_name = "PyLayer";
  type->tp_basicsize = sizeof(PyLayerObject);
  type->tp_dealloc = (destructor)PyLayerDealloc;
  type->tp_methods = pylayer_methods;
  type->tp_getset = pylayer_properties;
494
  type->tp_new = (newfunc)PyLayerNew;
W
wanghuancoder 已提交
495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522
  Py_INCREF(&PyBaseObject_Type);
  type->tp_base = reinterpret_cast<PyTypeObject*>(&PyBaseObject_Type);
  type->tp_flags |=
      Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
#if PY_VERSION_HEX >= 0x03050000
  type->tp_as_async = &heap_type->as_async;
#endif
  p_pylayer_type = type;

  if (PyType_Ready(type) < 0) {
    PADDLE_THROW(platform::errors::Fatal(
        "Init Paddle error in BindEager(PyType_Ready)."));
    return;
  }

  Py_INCREF(type);
  if (PyModule_AddObject(module, "PyLayer", reinterpret_cast<PyObject*>(type)) <
      0) {
    Py_DECREF(type);
    Py_DECREF(module);
    PADDLE_THROW(platform::errors::Fatal(
        "Init Paddle error in BindEager(PyModule_AddObject)."));
    return;
  }
}

}  // namespace pybind
}  // namespace paddle