diff --git a/paddle/fluid/operators/mul_op_npu.cc b/paddle/fluid/operators/mul_op_npu.cc new file mode 100644 index 0000000000000000000000000000000000000000..cf057cc339c621fcbd6d9f710f6b8c6983189600 --- /dev/null +++ b/paddle/fluid/operators/mul_op_npu.cc @@ -0,0 +1,243 @@ +/* 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 +#include + +#include "paddle/fluid/operators/mul_op.h" +#include "paddle/fluid/operators/npu_op_runner.h" + +namespace paddle { +namespace operators { + +template +class MulNPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* out = ctx.Output("Out"); + int x_num_col_dims = ctx.Attr("x_num_col_dims"); + int y_num_col_dims = ctx.Attr("y_num_col_dims"); + auto stream = + ctx.template device_context() + .stream(); + if (x_num_col_dims == 1 && y_num_col_dims == 1) { + if (x->dims().size() == 2 && y->dims().size() == 2) { + out->mutable_data(ctx.GetPlace()); + auto runner = + NpuOpRunner("MatMul", {*x, *y}, {*out}, + {{"transpose_x1", false}, {"transpose_x2", false}}); + + runner.Run(stream); + } else if (x->dims().size() == 3 && y->dims().size() == 2) { + // reshape + Tensor tmp_x(x->type()); + int64_t sec_dim = x->dims()[1] * x->dims()[2]; + int64_t first_dim = x->dims()[0]; + tmp_x.Resize(framework::make_ddim({first_dim, sec_dim})); + tmp_x.mutable_data(ctx.GetPlace()); + framework::TensorCopy( + *x, ctx.GetPlace(), + ctx.template device_context(), &tmp_x); + tmp_x.Resize(framework::make_ddim({first_dim, sec_dim})); + out->mutable_data(ctx.GetPlace()); + // matmul + auto runner = + NpuOpRunner("MatMul", {tmp_x, *y}, {*out}, + {{"transpose_x1", false}, {"transpose_x2", false}}); + runner.Run(stream); + } else { + PADDLE_THROW(platform::errors::InvalidArgument("not suppert dims")); + } + // to do other + } else if (x->dims().size() == 3 && y->dims().size() == 2) { + // for example: x.shape=[2, 3, 4] y.shape=[4, 5], expect [2, 3, 5] + PADDLE_ENFORCE_EQ(x_num_col_dims, 2, + platform::errors::InvalidArgument( + "now only support x_num_col_dims == 2: but got %d", + x_num_col_dims)); + // flatten => x.shape=[6, 4] + Tensor tmp_x(x->type()); + int64_t first_dim = x->dims()[0] * x->dims()[1]; + int64_t sec_dim = x->dims()[2]; + tmp_x.Resize(framework::make_ddim({first_dim, sec_dim})); + tmp_x.mutable_data(ctx.GetPlace()); + framework::TensorCopy( + *x, ctx.GetPlace(), + ctx.template device_context(), &tmp_x); + tmp_x.Resize(framework::make_ddim({first_dim, sec_dim})); + + // matmul [6,4] , [4, 5] => [6, 5] + Tensor tmp_matmul(x->type()); + tmp_matmul.Resize(framework::make_ddim({first_dim, y->dims()[1]})); + tmp_matmul.mutable_data(ctx.GetPlace()); + + auto runner_matmul = + NpuOpRunner("MatMul", {tmp_x, *y}, {tmp_matmul}, + {{"transpose_x1", false}, {"transpose_x2", false}}); + + runner_matmul.Run(stream); + // reshape [6, 5] => [2, 3, 5] + (*out).Resize( + framework::make_ddim({x->dims()[0], x->dims()[1], y->dims()[1]})); + out->mutable_data(ctx.GetPlace(), x->type()); + framework::TensorCopy( + tmp_matmul, ctx.GetPlace(), + ctx.template device_context(), out); + (*out).Resize( + framework::make_ddim({x->dims()[0], x->dims()[1], y->dims()[1]})); + } + } +}; + +template +class MulGradNPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* dout = ctx.Input(framework::GradVarName("Out")); + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dy = ctx.Output(framework::GradVarName("Y")); + int x_num_col_dims = ctx.Attr("x_num_col_dims"); + int y_num_col_dims = ctx.Attr("y_num_col_dims"); + auto stream = + ctx.template device_context() + .stream(); + if (x_num_col_dims == 1 && y_num_col_dims == 1) { + if (x->dims().size() == 2 && y->dims().size() == 2) { + if (dx) { + dx->mutable_data(ctx.GetPlace()); + auto runner_dx = + NpuOpRunner("MatMul", {*dout, *y}, {*dx}, + {{"transpose_x1", false}, {"transpose_x2", true}}); + + runner_dx.Run(stream); + } + + if (dy) { + dy->mutable_data(ctx.GetPlace()); + auto runner_dy = + NpuOpRunner("MatMul", {*x, *dout}, {*dy}, + {{"transpose_x1", true}, {"transpose_x2", false}}); + + runner_dy.Run(stream); + } + } else if (x->dims().size() == 3 && y->dims().size() == 2) { + // flatten => x.shape=[6, 4] + // matmul + if (dx) { + // matmul [2, 5] * [12, 5] => [2, 12] + Tensor tmp_matmul(y->type()); + tmp_matmul.Resize( + framework::make_ddim({dout->dims()[0], y->dims()[0]})); + tmp_matmul.mutable_data(ctx.GetPlace()); + auto runner_matmul = + NpuOpRunner("MatMul", {*dout, *y}, {tmp_matmul}, + {{"transpose_x1", false}, {"transpose_x2", true}}); + runner_matmul.Run(stream); + // reshape [2, 12] => [2, 3, 4] + dx->mutable_data(ctx.GetPlace(), x->type()); + framework::TensorCopy( + tmp_matmul, ctx.GetPlace(), + ctx.template device_context(), dx); + } + + if (dy) { + // flatten + Tensor tmp_x(x->type()); + int64_t sec_dim = x->dims()[1] * x->dims()[2]; + int64_t first_dim = x->dims()[0]; + tmp_x.Resize(framework::make_ddim({first_dim, sec_dim})); + tmp_x.mutable_data(ctx.GetPlace()); + framework::TensorCopy( + *x, ctx.GetPlace(), + ctx.template device_context(), &tmp_x); + tmp_x.Resize(framework::make_ddim({first_dim, sec_dim})); + dy->mutable_data(ctx.GetPlace()); + auto runner_dy = + NpuOpRunner("MatMul", {tmp_x, *dout}, {*dy}, + {{"transpose_x1", true}, {"transpose_x2", false}}); + + runner_dy.Run(stream); + } + } + } else if (x->dims().size() == 3 && y->dims().size() == 2) { + // for example: x.shape=[2, 3, 4] y.shape=[4, 5], expect [2, 3, 5] + PADDLE_ENFORCE_EQ(x_num_col_dims, 2, + platform::errors::InvalidArgument( + "now only support x_num_col_dims == 2: but got %d", + x_num_col_dims)); + // tmp_dout both used by dx and dy + Tensor tmp_dout(x->type()); + int64_t dout_first_dim = dout->dims()[0] * dout->dims()[1]; + int64_t dout_sec_dim = dout->dims()[2]; + tmp_dout.Resize(framework::make_ddim({dout_first_dim, dout_sec_dim})); + tmp_dout.mutable_data(ctx.GetPlace()); + framework::TensorCopy( + *dout, ctx.GetPlace(), + ctx.template device_context(), &tmp_dout); + tmp_dout.Resize(framework::make_ddim({dout_first_dim, dout_sec_dim})); + + if (dx) { + // tmp_dout * y [6,5] * [4,5] => [6, 4] + Tensor tmp_matmul(y->type()); + tmp_matmul.Resize(framework::make_ddim({dout_first_dim, y->dims()[0]})); + tmp_matmul.mutable_data(ctx.GetPlace()); + auto runner_matmul = + NpuOpRunner("MatMul", {tmp_dout, *y}, {tmp_matmul}, + {{"transpose_x1", false}, {"transpose_x2", true}}); + runner_matmul.Run(stream); + // reshape [6,4] => [2, 3, 4] + dx->mutable_data(ctx.GetPlace(), x->type()); + framework::TensorCopy( + tmp_matmul, ctx.GetPlace(), + ctx.template device_context(), dx); + } + if (dy) { + // flatten x.shape [2,3,4] => [6, 4] + Tensor tmp_x(x->type()); + int64_t first_dim = x->dims()[0] * x->dims()[1]; + int64_t sec_dim = x->dims()[2]; + tmp_x.Resize(framework::make_ddim({first_dim, sec_dim})); + tmp_x.mutable_data(ctx.GetPlace()); + framework::TensorCopy( + *x, ctx.GetPlace(), + ctx.template device_context(), &tmp_x); + tmp_x.Resize(framework::make_ddim({first_dim, sec_dim})); + // mamtul [6,4] [6,5] =>[4,5] + dy->mutable_data(ctx.GetPlace()); + auto runner_dy = + NpuOpRunner("MatMul", {tmp_x, tmp_dout}, {*dy}, + {{"transpose_x1", true}, {"transpose_x2", false}}); + runner_dy.Run(stream); + } + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_NPU_KERNEL( + mul, ops::MulNPUKernel, + ops::MulNPUKernel); +REGISTER_OP_NPU_KERNEL( + mul_grad, ops::MulGradNPUKernel, + ops::MulGradNPUKernel); diff --git a/python/paddle/fluid/tests/unittests/npu/test_mul_op_npu.py b/python/paddle/fluid/tests/unittests/npu/test_mul_op_npu.py new file mode 100644 index 0000000000000000000000000000000000000000..e65a3dac73928cd48c43e0d6eb4ebcc2a84e9d2d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/npu/test_mul_op_npu.py @@ -0,0 +1,326 @@ +# 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. + +from __future__ import print_function + +import numpy as np +import unittest +import sys +sys.path.append("..") +from op_test import OpTest +import paddle +import paddle.fluid as fluid + +paddle.enable_static() +SEED = 2021 + + +class TestMul(OpTest): + def config(self): + self.x_shape = (32, 5) + self.y_shape = (5, 100) + + def setUp(self): + self.set_npu() + self.op_type = "mul" + self.place = paddle.NPUPlace(0) + self.init_dtype() + self.config() + np.random.seed(SEED) + self.inputs = { + 'X': np.random.random(self.x_shape).astype(self.dtype), + 'Y': np.random.random(self.y_shape).astype(self.dtype) + } + self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} + + def set_npu(self): + self.__class__.use_npu = True + self.__class__.no_need_check_grad = True + + def init_dtype(self): + self.dtype = np.float32 + + def test_check_output(self): + self.check_output_with_place(self.place, check_dygraph=False, atol=1e-5) + + + # +class TestMulFP16(TestMul): + """ + case 2 + """ + + def init_dtype(self): + self.dtype = np.float16 + + +class TestMul3(TestMul): + """ + case 3 + """ + + def config(self): + self.x_shape = (2, 2, 5) + self.y_shape = (10, 5) + + def setUp(self): + self.set_npu() + self.op_type = "mul" + self.place = paddle.NPUPlace(0) + self.init_dtype() + self.config() + np.random.seed(SEED) + self.inputs = { + 'X': np.random.random(self.x_shape).astype(self.dtype), + 'Y': np.random.random(self.y_shape).astype(self.dtype) + } + self.outputs = { + 'Out': np.dot(self.inputs['X'].reshape(2, 10), self.inputs['Y']) + } + + +class TestMul4(TestMul): + """ + case 4 + """ + + def config(self): + self.x_shape = (2, 3, 4) + self.y_shape = (4, 5) + + def setUp(self): + self.set_npu() + self.op_type = "mul" + self.place = paddle.NPUPlace(0) + self.init_dtype() + self.config() + np.random.seed(SEED) + self.inputs = { + 'X': np.random.random(self.x_shape).astype(self.dtype), + 'Y': np.random.random(self.y_shape).astype(self.dtype) + } + self.attrs = {"x_num_col_dims": 2} + self.outputs = {'Out': np.matmul(self.inputs['X'], self.inputs['Y'])} + + +@unittest.skipIf(not paddle.is_compiled_with_npu(), + "core is not compiled with NPU") +class TestMulNet(unittest.TestCase): + def _test(self, run_npu=True): + main_prog = paddle.static.Program() + startup_prog = paddle.static.Program() + main_prog.random_seed = SEED + startup_prog.random_seed = SEED + np.random.seed(SEED) + + a_np = np.random.random(size=(2, 3)).astype('float32') + b_np = np.random.random(size=(2, 3)).astype('float32') + c_np = np.random.random(size=(3, 2)).astype('float32') + d_np = np.random.random(size=(3, 2)).astype('float32') + label_np = np.random.randint(2, size=(2, 1)).astype('int64') + + with paddle.static.program_guard(main_prog, startup_prog): + a = paddle.static.data(name="a", shape=[2, 3], dtype='float32') + b = paddle.static.data(name="b", shape=[2, 3], dtype='float32') + c = paddle.static.data(name="c", shape=[3, 2], dtype='float32') + d = paddle.static.data(name="d", shape=[3, 2], dtype='float32') + label = paddle.static.data( + name="label", shape=[2, 1], dtype='int64') + + sum_1 = paddle.add(a, b) + sum_2 = paddle.add(c, d) + result = paddle.fluid.layers.mul(sum_1, sum_2) + + fc_1 = fluid.layers.fc(input=result, size=8) + prediction = fluid.layers.fc(input=fc_1, size=2, act='softmax') + + cost = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.reduce_mean(cost) + sgd = fluid.optimizer.SGD(learning_rate=0.01) + sgd.minimize(loss) + + if run_npu: + place = paddle.NPUPlace(0) + else: + place = paddle.CPUPlace() + exe = paddle.static.Executor(place) + exe.run(startup_prog) + + print("TestMulNet Start run on {} . ".format(place)) + for epoch in range(100): + + pred_res, loss_res = exe.run(main_prog, + feed={ + "a": a_np, + "b": b_np, + "c": c_np, + "d": d_np, + "label": label_np + }, + fetch_list=[prediction, loss]) + if epoch % 10 == 0: + print("Epoch {} | Prediction[0]: {}, Loss: {}".format( + epoch, pred_res[0], loss_res)) + + return pred_res, loss_res + + def test_npu(self): + cpu_pred, cpu_loss = self._test(False) + npu_pred, npu_loss = self._test(True) + + self.assertTrue(np.allclose(npu_pred, cpu_pred)) + self.assertTrue(np.allclose(npu_loss, cpu_loss)) + + +@unittest.skipIf(not paddle.is_compiled_with_npu(), + "core is not compiled with NPU") +class TestMulNet3_2(unittest.TestCase): + def _test(self, run_npu=True): + main_prog = paddle.static.Program() + startup_prog = paddle.static.Program() + main_prog.random_seed = SEED + startup_prog.random_seed = SEED + np.random.seed(SEED) + + a_np = np.random.random(size=(2, 3, 4)).astype('float32') + b_np = np.random.random(size=(2, 3, 4)).astype('float32') + c_np = np.random.random(size=(12, 5)).astype('float32') + d_np = np.random.random(size=(12, 5)).astype('float32') + label_np = np.random.randint(2, size=(2, 1)).astype('int64') + + with paddle.static.program_guard(main_prog, startup_prog): + a = paddle.static.data(name="a", shape=[2, 3, 4], dtype='float32') + b = paddle.static.data(name="b", shape=[2, 3, 4], dtype='float32') + c = paddle.static.data(name="c", shape=[12, 5], dtype='float32') + d = paddle.static.data(name="d", shape=[12, 5], dtype='float32') + label = paddle.static.data( + name="label", shape=[2, 1], dtype='int64') + + sum_1 = paddle.add(a, b) + sum_2 = paddle.add(c, d) + result = paddle.fluid.layers.mul(sum_1, sum_2) + + fc_1 = fluid.layers.fc(input=result, size=8) + prediction = fluid.layers.fc(input=fc_1, size=2, act='softmax') + + cost = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.reduce_mean(cost) + sgd = fluid.optimizer.SGD(learning_rate=0.01) + sgd.minimize(loss) + + if run_npu: + place = paddle.NPUPlace(0) + else: + place = paddle.CPUPlace() + exe = paddle.static.Executor(place) + exe.run(startup_prog) + + print("testMulNet3_2 tart run on {}".format(place)) + for epoch in range(100): + + pred_res, loss_res = exe.run(main_prog, + feed={ + "a": a_np, + "b": b_np, + "c": c_np, + "d": d_np, + "label": label_np + }, + fetch_list=[prediction, loss]) + if epoch % 10 == 0: + print("Epoch {} | Prediction[0]: {}, Loss: {}".format( + epoch, pred_res[0], loss_res)) + + return pred_res, loss_res + + def test_npu(self): + cpu_pred, cpu_loss = self._test(False) + npu_pred, npu_loss = self._test(True) + + self.assertTrue(np.allclose(npu_pred, cpu_pred)) + self.assertTrue(np.allclose(npu_loss, cpu_loss)) + + +@unittest.skipIf(not paddle.is_compiled_with_npu(), + "core is not compiled with NPU") +class TestMulNet3_2_xc2(unittest.TestCase): + def _test(self, run_npu=True): + main_prog = paddle.static.Program() + startup_prog = paddle.static.Program() + main_prog.random_seed = SEED + startup_prog.random_seed = SEED + np.random.seed(SEED) + + a_np = np.random.random(size=(2, 3, 4)).astype('float32') + b_np = np.random.random(size=(2, 3, 4)).astype('float32') + c_np = np.random.random(size=(4, 5)).astype('float32') + d_np = np.random.random(size=(4, 5)).astype('float32') + label_np = np.random.randint(2, size=(2, 1)).astype('int64') + + with paddle.static.program_guard(main_prog, startup_prog): + a = paddle.static.data(name="a", shape=[2, 3, 4], dtype='float32') + b = paddle.static.data(name="b", shape=[2, 3, 4], dtype='float32') + c = paddle.static.data(name="c", shape=[4, 5], dtype='float32') + d = paddle.static.data(name="d", shape=[4, 5], dtype='float32') + label = paddle.static.data( + name="label", shape=[2, 1], dtype='int64') + + sum_1 = paddle.add(a, b) + sum_2 = paddle.add(c, d) + result = paddle.fluid.layers.mul(sum_1, sum_2, x_num_col_dims=2) + result_re = paddle.reshape(result, shape=[2, 15]) + + fc_1 = fluid.layers.fc(input=result_re, size=8) + prediction = fluid.layers.fc(input=fc_1, size=2, act='softmax') + + cost = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.reduce_mean(cost) + sgd = fluid.optimizer.SGD(learning_rate=0.01) + sgd.minimize(loss) + + if run_npu: + place = paddle.NPUPlace(0) + else: + place = paddle.CPUPlace() + exe = paddle.static.Executor(place) + exe.run(startup_prog) + + print("TestMulNet3_2_xc2. Start run on {}".format(place)) + for epoch in range(100): + + pred_res, loss_res = exe.run(main_prog, + feed={ + "a": a_np, + "b": b_np, + "c": c_np, + "d": d_np, + "label": label_np + }, + fetch_list=[prediction, loss]) + if epoch % 10 == 0: + print("Epoch {} | Prediction[0]: {}, Loss: {}".format( + epoch, pred_res[0], loss_res)) + + return pred_res, loss_res + + def test_npu(self): + cpu_pred, cpu_loss = self._test(False) + npu_pred, npu_loss = self._test(True) + + self.assertTrue(np.allclose(npu_pred, cpu_pred)) + self.assertTrue(np.allclose(npu_loss, cpu_loss)) + + +if __name__ == '__main__': + unittest.main()