未验证 提交 781f4028 编写于 作者: L Leo Chen 提交者: GitHub

copy found_inf to cpu in advance to improve performance (#34274)

* copy found_inf to cpu in advance to improve performance

* add npu test

* add npu test

* refine code

* refine memcpy op

* fix adam
上级 c342651e
develop Ligoml-patch-1 ZHUI-patch-1 add_some_yaml_config addfile ascendrelease cherry_undefined_var delete_delete_addfile delete_disable_iterable_dataset_unittest delete_fix_retry_ci delete_fix_undefined_var delete_improve_sccache delete_paralleltest delete_prv-disable-more-cache delete_revert-34910-spinlocks_for_allocator delete_revert-35069-revert-34910-spinlocks_for_allocator delete_revert-36057-dev/read_flags_in_ut dingjiaweiww-patch-1 disable_iterable_dataset_unittest dy2static enable_eager_model_test final_state_gen_python_c final_state_intermediate fix-numpy-issue fix_concat_slice fix_npu_ci fix_op_flops fix_retry_ci fix_rnn_docs fix_tensor_type fix_undefined_var fixiscan fixiscan1 fixiscan2 fixiscan3 improve_sccache incubate/infrt inplace_addto make_flag_adding_easier move_embedding_to_phi move_histogram_to_pten move_sgd_to_phi move_slice_to_pten move_temporal_shift_to_phi move_yolo_box_to_phi npu_fix_alloc paralleltest preln_ernie prv-disable-more-cache prv-md-even-more prv-onednn-2.5 pten_tensor_refactor release/2.2 release/2.3 release/2.3-fc-ernie-fix release/2.4 revert-34406-add_copy_from_tensor revert-34910-spinlocks_for_allocator revert-35069-revert-34910-spinlocks_for_allocator revert-36057-dev/read_flags_in_ut revert-36201-refine_fast_threaded_ssa_graph_executor revert-36985-add_license revert-37318-refactor_dygraph_to_eager revert-37926-eager_coreops_500 revert-37956-revert-37727-pylayer_support_tuple revert-38100-mingdong revert-38301-allocation_rearrange_pr revert-38703-numpy_bf16_package_reupload revert-38732-remove_useless_header_in_elementwise_mul_grad revert-38959-Reduce_Grad revert-39143-adjust_empty revert-39227-move_trace_op_to_pten revert-39268-dev/remove_concat_fluid_kernel revert-40170-support_partial_grad revert-41056-revert-40727-move_some_activaion_to_phi revert-41065-revert-40993-mv_ele_floordiv_pow revert-41068-revert-40790-phi_new revert-41944-smaller_inference_api_test revert-42149-do-not-reset-default-stream-for-stream-safe-cuda-allocator revert-43155-fix_ut_tempfile revert-43882-revert-41944-smaller_inference_api_test revert-45808-phi/simplify_size_op revert-46827-deform_comment support_weight_transpose zhiqiu-patch-1 v2.4.0-rc0 v2.3.2 v2.3.1 v2.3.0 v2.3.0-rc0 v2.2.2 v2.2.1 v2.2.0 v2.2.0-rc0 v2.2.0-bak0
无相关合并请求
/* Copyright (c) 2021 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. */
#include <memory>
#include <string>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/op_version_registry.h"
#include "paddle/fluid/operators/npu_op_runner.h"
namespace paddle {
namespace operators {
template <typename DeviceContext, typename T>
class MatMulNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<framework::Tensor>("X");
auto* y = ctx.Input<framework::Tensor>("Y");
auto* out = ctx.Output<framework::Tensor>("Out");
bool transpose_x = ctx.Attr<bool>("transpose_X");
bool transpose_y = ctx.Attr<bool>("transpose_Y");
if (x->dims().size() == 2) {
out->mutable_data<T>(ctx.GetPlace());
const auto& runner = NpuOpRunner(
"MatMul", {*x, *y}, {*out},
{{"transpose_x1", transpose_x}, {"transpose_x2", transpose_y}});
auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();
runner.Run(stream);
} else if (x->dims().size() > 2) {
out->mutable_data<T>(ctx.GetPlace());
const auto& runner =
NpuOpRunner("BatchMatMul", {*x, *y}, {*out},
{{"adj_x1", transpose_x}, {"adj_x2", transpose_y}});
auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();
runner.Run(stream);
}
}
};
template <typename DeviceContext, typename T>
class MatMulGradNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<framework::Tensor>("X");
auto* y = ctx.Input<framework::Tensor>("Y");
auto* dout = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto* dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
bool transpose_y = ctx.Attr<bool>("transpose_Y");
auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();
if (x->dims().size() == 2) {
if (transpose_y) {
if (dx) {
dx->mutable_data<T>(ctx.GetPlace());
const auto& runner_dx =
NpuOpRunner("MatMul", {*dout, *y}, {*dx},
{{"transpose_x1", false}, {"transpose_x2", false}});
runner_dx.Run(stream);
}
if (dy) {
dy->mutable_data<T>(ctx.GetPlace());
const auto& runner_dy =
NpuOpRunner("MatMul", {*dout, *x}, {*dy},
{{"transpose_x1", true}, {"transpose_x2", false}});
runner_dy.Run(stream);
}
} else {
if (dx) {
dx->mutable_data<T>(ctx.GetPlace());
const auto& runner_dx =
NpuOpRunner("MatMul", {*dout, *y}, {*dx},
{{"transpose_x1", false}, {"transpose_x2", true}});
runner_dx.Run(stream);
}
if (dy) {
dy->mutable_data<T>(ctx.GetPlace());
const auto& runner_dy =
NpuOpRunner("MatMul", {*x, *dout}, {*dy},
{{"transpose_x1", true}, {"transpose_x2", false}});
runner_dy.Run(stream);
}
}
} else if (x->dims().size() > 2) {
if (transpose_y) {
if (dx) {
dx->mutable_data<T>(ctx.GetPlace());
const auto& runner_dx =
NpuOpRunner("BatchMatMul", {*dout, *y}, {*dx},
{{"adj_x1", false}, {"adj_x2", false}});
runner_dx.Run(stream);
}
if (dy) {
dy->mutable_data<T>(ctx.GetPlace());
const auto& runner_dy =
NpuOpRunner("BatchMatMul", {*dout, *x}, {*dy},
{{"adj_x1", true}, {"adj_x2", false}});
runner_dy.Run(stream);
}
} else {
if (dx) {
dx->mutable_data<T>(ctx.GetPlace());
const auto& runner_dx =
NpuOpRunner("BatchMatMul", {*dout, *y}, {*dx},
{{"adj_x1", false}, {"adj_x2", true}});
runner_dx.Run(stream);
}
if (dy) {
dy->mutable_data<T>(ctx.GetPlace());
if ((x->dims().size() == 3) && (dout->dims().size() == 3) &&
(dy->dims().size() == 2)) {
framework::Tensor dout_;
dout_.ShareDataWith(*dout);
std::vector<int> vec_dim = framework::vectorize<int>(dout_.dims());
std::vector<int> vec_dim_v{vec_dim[0] * vec_dim[1], vec_dim[2]};
dout_.Resize(framework::make_ddim(vec_dim_v));
framework::Tensor x_;
x_.ShareDataWith(*x);
std::vector<int> vec_dim_x = framework::vectorize<int>(x_.dims());
std::vector<int> vec_dim_x_v{vec_dim_x[0] * vec_dim_x[1],
vec_dim_x[2]};
x_.Resize(framework::make_ddim(vec_dim_x_v));
const auto& runner_dy =
NpuOpRunner("MatMul", {x_, dout_}, {*dy},
{{"transpose_x1", true}, {"transpose_x2", false}});
runner_dy.Run(stream);
} else {
const auto& runner_dy =
NpuOpRunner("BatchMatMul", {*x, *dout}, {*dy},
{{"adj_x1", true}, {"adj_x2", false}});
runner_dy.Run(stream);
}
}
}
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_NPU_KERNEL(
matmul, ops::MatMulNPUKernel<paddle::platform::NPUDeviceContext, float>,
ops::MatMulNPUKernel<paddle::platform::NPUDeviceContext,
paddle::platform::float16>);
REGISTER_OP_NPU_KERNEL(
matmul_grad,
ops::MatMulGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
ops::MatMulGradNPUKernel<paddle::platform::NPUDeviceContext,
paddle::platform::float16>);
......@@ -51,17 +51,14 @@ class MemcpyFunctor {
} else if (dst_place_type_ == 1) {
framework::TensorCopy(lod_tensor, dev_ctx_.GetPlace(), dev_ctx_,
&out_tensor);
}
} else if (dst_place_type_ == 0) {
framework::TensorCopySync(lod_tensor, platform::CPUPlace(), &out_tensor);
#ifdef PADDLE_WITH_ASCEND_CL
else if (dst_place_type_ == 0) { // NOLINT
framework::TensorCopy(lod_tensor, platform::CPUPlace(), dev_ctx_,
&out_tensor);
} else if (dst_place_type_ == 4) {
framework::TensorCopy(lod_tensor, dev_ctx_.GetPlace(), dev_ctx_,
&out_tensor);
}
#endif
else { // NOLINT
} else {
PADDLE_THROW(platform::errors::Unimplemented(
"memcpy dst_place_type: %d is not supported yet.", dst_place_type_));
}
......
......@@ -122,7 +122,8 @@ framework::OpKernelType AdamOp::GetExpectedKernelType(
framework::OpKernelType AdamOp::GetKernelTypeForVar(
const std::string &var_name, const framework::Tensor &tensor,
const framework::OpKernelType &expected_kernel_type) const {
if (var_name == "Beta1Pow" || var_name == "Beta2Pow") {
if (var_name == "Beta1Pow" || var_name == "Beta2Pow" ||
var_name == "SkipUpdate") {
return expected_kernel_type;
} else {
return framework::OpKernelType(expected_kernel_type.data_type_,
......
......@@ -141,7 +141,7 @@ class AdamNPUKernel : public framework::OpKernel<T> {
if (ctx.HasInput("Beta2Tensor")) {
beta2_tensor = ctx.Input<framework::Tensor>("Beta2Tensor");
PADDLE_ENFORCE_EQ(beta1_tensor->numel(), 1,
PADDLE_ENFORCE_EQ(beta2_tensor->numel(), 1,
platform::errors::InvalidArgument(
"Input(Beta2Tensor) size must be 1, but get %d",
beta2_tensor->numel()));
......
......@@ -400,6 +400,10 @@ class OptimizerWithMixedPrecision(object):
name="update_loss_scaling")
# Pass found_inf to adam, to skip update for not only param, but also momentum and beta_pow
if isinstance(self._optimizer, paddle.fluid.optimizer.Adam):
# NOTE(zhiqiu): Since found_inf needs to be on cpu in adam op, we
# copy it in advance to avoid multiple time copies.
found_inf = paddle.tensor.creation._memcpy(found_inf,
paddle.CPUPlace())
self._optimizer._set_auxiliary_var('found_inf', found_inf)
optimize_ops = self._optimizer.apply_gradients(params_grads)
return optimize_ops
......
# Copyright (c) 2020 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.
import unittest
import sys
import paddle
sys.path.append("..")
import test_mixed_precision
paddle.enable_static()
class AMPTestNpu(test_mixed_precision.AMPTest):
def setUp(self):
self.place = paddle.NPUPlace(0)
if __name__ == '__main__':
unittest.main()
......@@ -144,32 +144,6 @@ class TestMemcpyOPError(unittest.TestCase):
feed={},
fetch_list=[selected_row_var.name, pinned_var.name])
def test_OTHER_PLACE_NotImplementedError(self):
main_program, pinned_var = self.get_prog()
lod_tensor_var = main_program.global_block().create_var( \
name="lod_tensor_0", dtype="float32", persistable=False, stop_gradient=True)
main_program.global_block().append_op(
type="fill_constant",
outputs={"Out": lod_tensor_var},
attrs={
"shape": lod_tensor_var.shape,
"dtype": lod_tensor_var.dtype,
"value": 1.0,
"place_type": 0
})
main_program.global_block().append_op(
type='memcpy',
inputs={'X': pinned_var},
outputs={'Out': lod_tensor_var},
attrs={'dst_place_type': 0, })
with self.assertRaises(NotImplementedError):
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
lod_tensor_var_, pinned_ = exe.run(
main_program,
feed={},
fetch_list=[lod_tensor_var.name, pinned_var.name])
class TestMemcpyApi(unittest.TestCase):
def test_api(self):
......
......@@ -47,6 +47,9 @@ class SimpleNet(nn.Layer):
class AMPTest(unittest.TestCase):
def setUp(self):
self.place = paddle.CUDAPlace(0)
def net(self):
input_size = 4096
output_size = 4096
......@@ -82,7 +85,8 @@ class AMPTest(unittest.TestCase):
fetch_list = [
loss, weight, moment1, beta_pow1, 'find_infinite_scale.tmp_0'
]
exe = paddle.static.Executor(paddle.CUDAPlace(0))
exe = paddle.static.Executor(self.place)
train_data = [
np.random.rand(batch_size, input_size).astype(np.float32)
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册
反馈
建议
客服 返回
顶部