Skip to content

P
Paddle

PaddlePaddle / Paddle
1 年多 前同步成功

通知 2305
Star 20932
Fork 5423
P
Paddle
未验证 提交 b451aff8 编写于 作者: J jjyaoao 提交者： GitHub
浏览文件
操作

delete paddle/fluid/operators/collective/*_npu.* (#52677)

上级 c1cad896
隐藏空白更改
内联 并排
Showing with 0 addition and 2688 deletion
paddle/fluid/operators/collective/c_allgather_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/c_allgather_op.h"
namespace paddle {
namespace operators {
template <typename T>
class CAllGatherOpASCENDKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"PaddlePaddle should compile with NPU."));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_allgather,
ops::CAllGatherOpASCENDKernel<int8_t>,
ops::CAllGatherOpASCENDKernel<int>,
ops::CAllGatherOpASCENDKernel<int64_t>,
ops::CAllGatherOpASCENDKernel<float>,
ops::CAllGatherOpASCENDKernel<plat::float16>);
paddle/fluid/operators/collective/c_allgather_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/c_allgather_op.h"
#include "paddle/fluid/operators/collective/c_allreduce_op.h"
#include "paddle/fluid/operators/collective/c_broadcast_op.h"
#include "paddle/fluid/operators/collective/c_reducescatter_op.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(c_allgather);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(c_allgather, NPU);
DECLARE_string(selected_npus);
template <typename T>
void PrintDebugInfo(const std::string preStr, const std::vector<T>& data) {
std::string debugstring = "";
for (auto ele : data) {
debugstring += std::to_string(ele) + std::string(",");
}
VLOG(2) << preStr << ":" << std::endl << debugstring;
}
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
void TestHCCLAllGatherOp(f::Scope* scope, const p::DeviceContext& ctx) {
// init
auto x = scope->Var("Data");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
std::vector<float> init;
int rank_id = atoi(getenv("RANK_ID"));
int num1 = 1;
int num2 = 4;
for (int64_t i = 0; i < num1 * num2; ++i) {
init.push_back(1.0 + rank_id);
}
PrintDebugInfo("input data", init);
paddle::framework::TensorFromVector(init, ctx, tensor_x);
tensor_x->Resize({num1, num2});
ctx.Wait();
auto place = ctx.GetPlace();
auto out = scope->Var("OutData");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
tensor_out->Resize({num1, num2});
tensor_out->mutable_data<float>(place); // allocate
ctx.Wait();
// run
f::AttributeMap attrs;
attrs["tag"] = std::string("tagx");
attrs["ring_id"] = 0;
attrs["nranks"] = 2;
auto op = f::OpRegistry::CreateOp(
"c_allgather", {{"X", {"Data"}}}, {{"Out", {"OutData"}}}, attrs);
for (int i = 0; i < 10; i++) {
op->Run(*scope, place);
}
ctx.Wait();
std::vector<float> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
ctx.Wait();
PrintDebugInfo("output data", out_vec);
EXPECT_EQ(out_vec.size(), init.size() * 2);
for (uint32_t i = 0; i < out_vec.size() / 2; i++) {
EXPECT_EQ(out_vec[i], 1.0);
}
for (uint32_t i = out_vec.size() / 2; i < out_vec.size(); i++) {
EXPECT_EQ(out_vec[i], 2.0);
}
}
TEST(c_allgather, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
// only support one device, if more than one device, use first default
p::NPUDeviceContext ctx(p::NPUPlace(atoi(FLAGS_selected_npus.c_str())));
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
TestHCCLAllGatherOp(&scope, ctx);
}
paddle/fluid/operators/collective/c_allreduce_max_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/c_allreduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(
c_allreduce_max,
ops::CAllReduceOpASCENDKernel<ops::kRedMax, int>,
ops::CAllReduceOpASCENDKernel<ops::kRedMax, int8_t>,
ops::CAllReduceOpASCENDKernel<ops::kRedMax, float>,
ops::CAllReduceOpASCENDKernel<ops::kRedMax, plat::float16>)
paddle/fluid/operators/collective/c_allreduce_max_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/c_allgather_op.h"
#include "paddle/fluid/operators/collective/c_allreduce_op.h"
#include "paddle/fluid/operators/collective/c_broadcast_op.h"
#include "paddle/fluid/operators/collective/c_reducescatter_op.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(c_allreduce_max);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(c_allreduce_max, NPU);
DECLARE_string(selected_npus);
template <typename T>
void PrintDebugInfo(const std::string preStr, const std::vector<T>& data) {
std::string debugstring = "";
for (auto ele : data) {
debugstring += std::to_string(ele) + std::string(",");
}
VLOG(2) << preStr << ":" << std::endl << debugstring;
}
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
void TestHCCLAllReduceOp(f::Scope* scope, const p::DeviceContext& ctx) {
// init
auto x = scope->Var("Data");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
std::vector<float> init;
int rank_id = atoi(getenv("RANK_ID"));
int num1 = 100;
int num2 = 100;
for (int64_t i = 0; i < num1 * num2; ++i) {
init.push_back(1.0 + rank_id * 3);
}
PrintDebugInfo("input data", init);
paddle::framework::TensorFromVector(init, ctx, tensor_x);
tensor_x->Resize({num1, num2});
ctx.Wait();
auto place = ctx.GetPlace();
auto out = scope->Var("OutData");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
tensor_out->Resize({num1, num2});
tensor_out->mutable_data<float>(place); // allocate
ctx.Wait();
// run
f::AttributeMap attrs;
attrs["tag"] = std::string("tagx");
attrs["ring_id"] = 0;
auto op = f::OpRegistry::CreateOp(
"c_allreduce_max", {{"X", {"Data"}}}, {{"Out", {"OutData"}}}, attrs);
for (int i = 0; i < 10; i++) {
op->Run(*scope, place);
}
ctx.Wait();
std::vector<float> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
ctx.Wait();
PrintDebugInfo("output data", out_vec);
EXPECT_EQ(out_vec.size(), init.size());
for (uint32_t i = 0; i < out_vec.size(); i++) {
EXPECT_EQ(out_vec[i], 4.0);
}
}
TEST(c_allreduce_max, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
// only support one device, if more than one device, use first default
p::NPUDeviceContext ctx(p::NPUPlace(atoi(FLAGS_selected_npus.c_str())));
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
TestHCCLAllReduceOp(&scope, ctx);
}
paddle/fluid/operators/collective/c_allreduce_min_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/c_allreduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(
c_allreduce_min,
ops::CAllReduceOpASCENDKernel<ops::kRedMin, int>,
ops::CAllReduceOpASCENDKernel<ops::kRedMin, int8_t>,
ops::CAllReduceOpASCENDKernel<ops::kRedMin, float>,
ops::CAllReduceOpASCENDKernel<ops::kRedMin, plat::float16>)
paddle/fluid/operators/collective/c_allreduce_prod_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/c_allreduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(
c_allreduce_prod,
ops::CAllReduceOpASCENDKernel<ops::kRedProd, int>,
ops::CAllReduceOpASCENDKernel<ops::kRedProd, int8_t>,
ops::CAllReduceOpASCENDKernel<ops::kRedProd, float>,
ops::CAllReduceOpASCENDKernel<ops::kRedProd, plat::float16>)
paddle/fluid/operators/collective/c_allreduce_sum_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/c_allreduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(
c_allreduce_sum,
ops::CAllReduceOpASCENDKernel<ops::kRedSum, int>,
ops::CAllReduceOpASCENDKernel<ops::kRedSum, int8_t>,
ops::CAllReduceOpASCENDKernel<ops::kRedSum, float>,
ops::CAllReduceOpASCENDKernel<ops::kRedSum, plat::float16>)
paddle/fluid/operators/collective/c_allreduce_sum_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/c_allreduce_op.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
// Node1: HCCL_WHITELIST_DISABLE=1 FLAGS_selected_npus=1 GLOG_v=4 RANK_ID=1
// DEVICE_ID=1 ./paddle/fluid/operators/collective/c_allreduce_sum_op_npu_test
// Node2: HCCL_WHITELIST_DISABLE=1 FLAGS_selected_npus=0 GLOG_v=4 RANK_ID=0
// DEVICE_ID=0 ./paddle/fluid/operators/collective/c_allreduce_sum_op_npu_test
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(c_allreduce_sum);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(c_allreduce_sum, NPU);
DECLARE_string(selected_npus);
template <typename T>
void PrintDebugInfo(const std::string preStr, const std::vector<T>& data) {
std::string debugstring = "";
std::cout << preStr << ":" << std::endl << debugstring;
for (auto ele : data) {
std::cout << ele << " ";
}
std::cout << std::endl;
}
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
template <typename T>
void TestHCCLAllReduceOp(f::Scope* scope,
const p::DeviceContext& ctx,
int iter) {
// init
auto x = scope->Var("Data");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
int rank_id = atoi(getenv("RANK_ID"));
int num1 = 3;
int num2 = 128;
std::vector<T> init;
for (int64_t i = 0; i < num1 * num2; ++i) {
init.push_back(static_cast<T>(1.0 + rank_id));
}
init[0] = static_cast<T>(std::numeric_limits<float>::quiet_NaN());
PrintDebugInfo("input data", init);
auto place = ctx.GetPlace();
paddle::framework::TensorFromVector(init, ctx, tensor_x);
tensor_x->Resize({num1, num2});
ctx.Wait();
auto out = scope->Var("OutData");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
tensor_out->Resize({num1, num2});
tensor_out->mutable_data<T>(place); // allocate
ctx.Wait();
// run
f::AttributeMap attrs;
attrs["tag"] = std::string("tagx_" + std::to_string(iter));
attrs["ring_id"] = 0;
attrs["use_calc_stream"] = 1;
auto op = f::OpRegistry::CreateOp(
"c_allreduce_sum", {{"X", {"Data"}}}, {{"Out", {"OutData"}}}, attrs);
for (int i = 0; i < 1; i++) {
op->Run(*scope, place);
}
ctx.Wait();
std::vector<T> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
ctx.Wait();
PrintDebugInfo("output data", out_vec);
float diff = static_cast<float>(out_vec[0]) - 65504;
EXPECT_TRUE(diff < 0.1 && diff > -0.1);
EXPECT_EQ(out_vec.size(), init.size());
for (uint32_t i = 1; i < 10; i++) {
EXPECT_EQ(out_vec[i], static_cast<paddle::platform::float16>(3.0));
}
}
TEST(c_allreduce_sum, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
p::NPUDeviceContext ctx(p::NPUPlace(atoi(FLAGS_selected_npus.c_str())));
// only support one device, if more than one device, use first default
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
TestHCCLAllReduceOp<paddle::platform::float16>(&scope, ctx, 1);
// TestHCCLAllReduceOp<float>(&scope, ctx, 0);
}
paddle/fluid/operators/collective/c_broadcast_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/c_broadcast_op.h"
namespace paddle {
namespace operators {
template <typename T>
class CBroadcastOpASCENDKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"PaddlePaddle should compile with NPU."));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_broadcast,
ops::CBroadcastOpASCENDKernel<int>,
ops::CBroadcastOpASCENDKernel<int8_t>,
ops::CBroadcastOpASCENDKernel<float>,
ops::CBroadcastOpASCENDKernel<plat::float16>);
paddle/fluid/operators/collective/c_broadcast_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/c_broadcast_op.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(c_broadcast);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(c_broadcast, NPU);
DECLARE_string(selected_npus);
template <typename T>
void PrintDebugInfo(const std::string preStr, const std::vector<T>& data) {
std::string debugstring = "";
for (auto ele : data) {
debugstring += std::to_string(ele) + std::string(",");
}
VLOG(2) << preStr << ":" << std::endl << debugstring;
}
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
void TestHCCLBroadcastOp(f::Scope* scope, const p::DeviceContext& ctx) {
// init
auto x = scope->Var("Data");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
int num = 2;
std::vector<float> init;
int rank_id = atoi(getenv("RANK_ID"));
for (int64_t i = 0; i < num * num; ++i) {
init.push_back(1.0 + rank_id);
}
PrintDebugInfo("input data", init);
paddle::framework::TensorFromVector(init, ctx, tensor_x);
tensor_x->Resize({num, num});
ctx.Wait();
auto place = ctx.GetPlace();
auto out = scope->Var("OutData");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
tensor_out->Resize({num, num});
tensor_out->mutable_data<float>(place); // allocate
ctx.Wait();
// run
f::AttributeMap attrs;
attrs["tag"] = std::string("tagx");
attrs["root"] = 0;
attrs["ring_id"] = 0;
auto op = f::OpRegistry::CreateOp(
"c_broadcast", {{"X", {"Data"}}}, {{"Out", {"OutData"}}}, attrs);
for (int i = 0; i < 10; i++) {
op->Run(*scope, place);
}
ctx.Wait();
std::vector<float> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
ctx.Wait();
PrintDebugInfo("output data", out_vec);
EXPECT_EQ(out_vec.size(), init.size());
for (uint32_t i = 0; i < out_vec.size(); i++) {
EXPECT_EQ(out_vec[i], 1.0);
}
}
TEST(c_broadcast, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
// only support one device, if more than one device, use first default
p::NPUDeviceContext ctx(p::NPUPlace(atoi(FLAGS_selected_npus.c_str())));
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
TestHCCLBroadcastOp(&scope, ctx);
}
paddle/fluid/operators/collective/c_embedding_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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/tensor_util.h"
#include "paddle/fluid/operators/collective/c_embedding_op.h"
namespace paddle {
namespace operators {
template <typename T>
inline void FillNPU(Tensor *dst,
T val,
const framework::ExecutionContext &context) {
Tensor value(dst->type());
value.mutable_data<T>({1}, context.GetPlace());
FillNpuTensorWithConstant<T>(&value, static_cast<T>(val));
auto stream =
context.template device_context<paddle::platform::NPUDeviceContext>()
.stream();
const auto &runner = NpuOpRunner(
"FillD", {value}, {*dst}, {{"dims", phi::vectorize(dst->dims())}});
runner.Run(stream);
}
template <typename T>
void shard_index(const Tensor &table_t,
const Tensor &ids_t,
int64_t start_idx,
const Tensor &id_t,
const framework::ExecutionContext &context) {
const int height = table_t.dims()[0];
auto stream =
context.template device_context<paddle::platform::NPUDeviceContext>()
.stream();
phi::DenseTensor id_t_d;
id_t_d.mutable_data<T>(ids_t.dims(), context.GetPlace());
FillNPU(&id_t_d, static_cast<T>(0.0), context);
id_t_d.Resize(ids_t.dims());
phi::DenseTensor id_t_u;
id_t_u.mutable_data<T>(ids_t.dims(), context.GetPlace());
FillNPU(&id_t_u, static_cast<T>(height - 1), context);
id_t_u.Resize(ids_t.dims());
phi::DenseTensor id_matched_d;
id_matched_d.mutable_data<bool>(ids_t.dims(), context.GetPlace());
phi::DenseTensor id_matched_u;
id_matched_u.mutable_data<bool>(ids_t.dims(), context.GetPlace());
phi::DenseTensor ignore_tensor;
ignore_tensor.mutable_data<T>(ids_t.dims(), context.GetPlace());
FillNPU(&ignore_tensor, static_cast<T>(height), context);
ignore_tensor.Resize(ids_t.dims());
NpuOpRunner sub_runner;
#if (CANN_VERSION_CODE >= 503003)
Tensor factor_tensor(ids_t.type());
factor_tensor.mutable_data<T>({1}, context.GetPlace());
paddle::framework::TensorFromVector(std::vector<T>{static_cast<T>(start_idx)},
context.device_context(),
&factor_tensor);
sub_runner.SetType("Sub")
.AddInput(ids_t)
.AddInput(factor_tensor)
.AddOutput(id_t);
#else
sub_runner.SetType("Sub")
.AddInput(ids_t)
.AddInput(std::vector<T>{static_cast<T>(start_idx)})
.AddOutput(id_t);
#endif
sub_runner.Run();
NpuOpRunner lessequal1_runner;
lessequal1_runner.SetType("LessEqual")
.AddInput(id_t)
.AddInput(id_t_u)
.AddOutput(id_matched_u);
lessequal1_runner.Run();
NpuOpRunner lessequal2_runner;
lessequal2_runner.SetType("LessEqual")
.AddInput(id_t_d)
.AddInput(id_t)
.AddOutput(id_matched_d);
lessequal2_runner.Run();
NpuOpRunner("Equal", {id_matched_u, id_matched_d}, {id_matched_d}, {})
.Run(stream);
NpuOpRunner("Select", {id_matched_d, id_t, ignore_tensor}, {id_t}, {})
.Run(stream);
}
template <typename TIds, typename T>
void NPUGetIdsEmbedding(const framework::ExecutionContext &context) {
auto *table_t = context.Input<phi::DenseTensor>("W");
auto *ids_t = context.Input<phi::DenseTensor>("Ids");
auto *output_t = context.Output<phi::DenseTensor>("Out");
const int64_t start_idx = context.Attr<int64_t>("start_index");
auto stream =
context.template device_context<paddle::platform::NPUDeviceContext>()
.stream();
phi::DenseTensor ids_t_local;
ids_t_local.mutable_data<TIds>(ids_t->dims(), context.GetPlace());
shard_index<TIds>(*table_t, *ids_t, start_idx, ids_t_local, context);
auto pad_shape = phi::make_ddim({table_t->dims()[0] + 1, table_t->dims()[1]});
phi::DenseTensor table_t_pad;
size_t mem_size = table_t->numel() * phi::SizeOf(table_t->dtype());
size_t line_mem_size = table_t->dims()[1] * phi::SizeOf(table_t->dtype());
PADDLE_ENFORCE_EQ(line_mem_size % 64,
0,
platform::errors::InvalidArgument(
"NPU only accept the second dim must align by 64"));
VLOG(10) << "mem_size:" << mem_size << ",line_mem_size:" << line_mem_size
<< ", pad_shape:" << pad_shape << ", table_dims:" << table_t->dims();
uint8_t *pad_data = reinterpret_cast<uint8_t *>(
table_t_pad.mutable_data<T>(pad_shape, context.GetPlace()));
platform::NPUMemcpyAsync(pad_data,
table_t->data<T>(),
mem_size,
ACL_MEMCPY_DEVICE_TO_DEVICE,
stream,
mem_size);
platform::NPUMemsetAsync(
pad_data + mem_size, 0, line_mem_size, stream, line_mem_size);
output_t->mutable_data<T>(context.GetPlace());
NpuOpRunner runner;
runner.SetType("GatherV2")
.AddInput(table_t_pad)
.AddInput(ids_t_local)
.AddInput(std::vector<int32_t>{0})
#if (CANN_VERSION_CODE >= 503003)
.AddAttrs({{"batch_dims", 0}})
#endif
.AddOutput(*output_t);
runner.Run();
}
template <typename T>
class CEmbeddingNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &context) const override {
auto *ids_t = context.Input<phi::DenseTensor>("Ids");
const auto &index_type = framework::TransToProtoVarType(ids_t->dtype());
if (index_type == framework::proto::VarType::INT32) {
NPUGetIdsEmbedding<int32_t, T>(context);
} else {
PADDLE_THROW(platform::errors::Unavailable(
"NPU c_embedding ids only support int32."));
}
}
};
template <typename TIds, typename T>
void NPUUpdateEmbedding(const framework::ExecutionContext &context) {
// get inputs
const int64_t start_idx = context.Attr<int64_t>("start_index");
auto ids_t = context.Input<phi::DenseTensor>("Ids");
auto d_output_t =
context.Input<phi::DenseTensor>(framework::GradVarName("Out"));
auto table_t = context.Input<phi::DenseTensor>("W");
auto table_grad_t =
context.Output<phi::DenseTensor>(framework::GradVarName("W"));
VLOG(10) << "ids_t:" << ids_t << ", d_output_t:" << d_output_t
<< ", table_t:" << table_t << ", table_grad_t" << table_grad_t;
auto stream =
context.template device_context<paddle::platform::NPUDeviceContext>()
.stream();
// convert ids_t to local valid
phi::DenseTensor ids_t_local;
ids_t_local.mutable_data<TIds>(ids_t->dims(), context.GetPlace());
shard_index<TIds>(*table_t, *ids_t, start_idx, ids_t_local, context);
// padding table_t -> table_t_pad
auto pad_shape = phi::make_ddim({table_t->dims()[0] + 1, table_t->dims()[1]});
phi::DenseTensor table_t_pad;
// set table_t_pad to zero
uint8_t *pad_data = reinterpret_cast<uint8_t *>(
table_t_pad.mutable_data<T>(pad_shape, context.GetPlace()));
size_t table_t_pad_mem_size =
table_t_pad.numel() *
framework::SizeOfType(
framework::TransToProtoVarType(table_t_pad.dtype()));
platform::NPUMemsetAsync(
pad_data, 0, table_t_pad_mem_size, stream, table_t_pad_mem_size);
// NOTE(zhiqiu): It seems in cann 20.1, the first input and output
// can be different tensor, but in cann 20.2+, it does inplace operation.
// Thus, the first input and output should be same tensor.
const auto &runner_scatter =
NpuOpRunner("ScatterAdd",
{table_t_pad, ids_t_local, *d_output_t},
{table_t_pad},
{{"use_locking", true}});
runner_scatter.Run(stream);
// copy table_t_pad to table_t
T *dst = table_grad_t->mutable_data<T>(table_t->dims(), context.GetPlace());
const size_t mem_size =
table_grad_t->numel() * phi::SizeOf(table_grad_t->dtype());
// check align
size_t line_mem_size =
table_grad_t->dims()[1] * phi::SizeOf(table_grad_t->dtype());
PADDLE_ENFORCE_EQ(line_mem_size % 64,
0,
platform::errors::InvalidArgument(
"NPU only accept the second dim must align by 64"));
platform::NPUMemcpyAsync(
dst, pad_data, mem_size, ACL_MEMCPY_DEVICE_TO_DEVICE, stream, mem_size);
}
template <typename T>
class CEmbeddingGradNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &context) const override {
auto *ids_t = context.Input<phi::DenseTensor>("Ids");
const auto &index_type = framework::TransToProtoVarType(ids_t->dtype());
if (index_type == framework::proto::VarType::INT32) {
NPUUpdateEmbedding<int32_t, T>(context);
} else {
PADDLE_THROW(
platform::errors::Unavailable("c_embedding ids only support int32."));
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_embedding,
ops::CEmbeddingNPUKernel<float>,
ops::CEmbeddingNPUKernel<double>,
ops::CEmbeddingNPUKernel<plat::float16>);
REGISTER_OP_NPU_KERNEL(c_embedding_grad,
ops::CEmbeddingGradNPUKernel<float>,
ops::CEmbeddingGradNPUKernel<double>,
ops::CEmbeddingGradNPUKernel<plat::float16>);
paddle/fluid/operators/collective/c_identity_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/c_identity_op.h"
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_identity,
ops::CIdentityOpKernel<float, plat::NPUPlace>,
ops::CIdentityOpKernel<double, plat::NPUPlace>,
ops::CIdentityOpKernel<int, plat::NPUPlace>,
ops::CIdentityOpKernel<int64_t, plat::NPUPlace>,
ops::CIdentityOpKernel<plat::float16, plat::NPUPlace>);
paddle/fluid/operators/collective/c_reduce_max_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* Copyright (c) 2019 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 "paddle/fluid/operators/collective/c_reduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_reduce_max,
ops::CReduceOpASCENDKernel<ops::kRedMax, int>,
ops::CReduceOpASCENDKernel<ops::kRedMax, int8_t>,
ops::CReduceOpASCENDKernel<ops::kRedMax, float>,
ops::CReduceOpASCENDKernel<ops::kRedMax, plat::float16>)
paddle/fluid/operators/collective/c_reduce_min_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* Copyright (c) 2019 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 "paddle/fluid/operators/collective/c_reduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_reduce_min,
ops::CReduceOpASCENDKernel<ops::kRedMin, int>,
ops::CReduceOpASCENDKernel<ops::kRedMin, int8_t>,
ops::CReduceOpASCENDKernel<ops::kRedMin, float>,
ops::CReduceOpASCENDKernel<ops::kRedMin, plat::float16>)
paddle/fluid/operators/collective/c_reduce_prod_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* Copyright (c) 2019 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 "paddle/fluid/operators/collective/c_reduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_reduce_prod,
ops::CReduceOpASCENDKernel<ops::kRedProd, int>,
ops::CReduceOpASCENDKernel<ops::kRedProd, int8_t>,
ops::CReduceOpASCENDKernel<ops::kRedProd, float>,
ops::CReduceOpASCENDKernel<ops::kRedProd, plat::float16>)
paddle/fluid/operators/collective/c_reduce_sum_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* Copyright (c) 2019 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 "paddle/fluid/operators/collective/c_reduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_reduce_sum,
ops::CReduceOpASCENDKernel<ops::kRedSum, int>,
ops::CReduceOpASCENDKernel<ops::kRedSum, int8_t>,
ops::CReduceOpASCENDKernel<ops::kRedSum, float>,
ops::CReduceOpASCENDKernel<ops::kRedSum, plat::float16>)
paddle/fluid/operators/collective/c_reduce_sum_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/c_reduce_op.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(c_reduce_sum);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(c_reduce_sum, NPU);
DECLARE_string(selected_npus);
template <typename T>
void PrintDebugInfo(const std::string preStr, const std::vector<T>& data) {
std::string debugstring = "";
for (auto ele : data) {
debugstring += std::to_string(ele) + std::string(",");
}
VLOG(3) << preStr << ":" << std::endl << debugstring;
}
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
void TestHCCLReduceOp(f::Scope* scope, const p::DeviceContext& ctx, int iter) {
// init
auto x = scope->Var("Data");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
int rank_id = atoi(getenv("RANK_ID"));
int num1 = 3;
int num2 = 128;
std::vector<float> init;
for (int64_t i = 0; i < num1 * num2; ++i) {
init.push_back(1.0 + rank_id);
}
PrintDebugInfo("input data", init);
auto place = ctx.GetPlace();
paddle::framework::TensorFromVector(init, ctx, tensor_x);
tensor_x->Resize({num1, num2});
ctx.Wait();
auto out = scope->Var("OutData");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
tensor_out->Resize({num1, num2});
tensor_out->mutable_data<float>(place); // allocate
ctx.Wait();
// run
f::AttributeMap attrs;
attrs["tag"] = std::string("tagx_" + std::to_string(iter));
attrs["ring_id"] = 0;
int root_id = 0;
attrs["root_id"] = root_id;
auto op = f::OpRegistry::CreateOp(
"c_reduce_sum", {{"X", {"Data"}}}, {{"Out", {"OutData"}}}, attrs);
op->Run(*scope, place);
ctx.Wait();
std::vector<float> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
ctx.Wait();
PrintDebugInfo("output data", out_vec);
EXPECT_EQ(out_vec.size(), init.size());
for (uint32_t i = 0; i < out_vec.size(); i++) {
if (rank_id == root_id) {
EXPECT_EQ(out_vec[i], 3.0);
} else {
EXPECT_EQ(out_vec[i], init[i]);
}
}
}
TEST(c_reduce_sum, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
// only support one device, if more than one device, use first default
p::NPUDeviceContext ctx(p::NPUPlace(atoi(FLAGS_selected_npus.c_str())));
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
for (int i = 0; i < 2; i++) {
VLOG(2) << "iter num: " << i;
TestHCCLReduceOp(&scope, ctx, i);
}
}
paddle/fluid/operators/collective/c_reducescatter_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/c_reducescatter_op.h"
namespace paddle {
namespace operators {
template <typename T>
class CReduceScatterOpAscendKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"PaddlePaddle should compile with NPU."));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(c_reducescatter,
ops::CReduceScatterOpAscendKernel<int8_t>,
ops::CReduceScatterOpAscendKernel<int>,
ops::CReduceScatterOpAscendKernel<float>,
ops::CReduceScatterOpAscendKernel<plat::float16>);
paddle/fluid/operators/collective/c_reducescatter_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/c_allgather_op.h"
#include "paddle/fluid/operators/collective/c_allreduce_op.h"
#include "paddle/fluid/operators/collective/c_broadcast_op.h"
#include "paddle/fluid/operators/collective/c_reducescatter_op.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(c_reducescatter);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(c_reducescatter, NPU);
DECLARE_string(selected_npus);
template <typename T>
void PrintDebugInfo(const std::string preStr, const std::vector<T>& data) {
std::string debugstring = "";
for (auto ele : data) {
debugstring += std::to_string(ele) + std::string(",");
}
VLOG(2) << preStr << ":" << std::endl << debugstring;
}
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
void TestHCCLReduceScatterOp(f::Scope* scope, const p::DeviceContext& ctx) {
// init
auto x = scope->Var("Data");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
std::vector<float> init;
int num1 = 4;
int num2 = 1;
for (int64_t i = 0; i < num1 * num2; ++i) {
init.push_back(1.0);
}
PrintDebugInfo("input data", init);
paddle::framework::TensorFromVector(init, ctx, tensor_x);
tensor_x->Resize({num1, num2});
ctx.Wait();
auto place = ctx.GetPlace();
auto out = scope->Var("OutData");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
tensor_out->Resize({num1, num2});
tensor_out->mutable_data<float>(place); // allocate
ctx.Wait();
// run
f::AttributeMap attrs;
attrs["tag"] = std::string("tagx");
attrs["ring_id"] = 0;
attrs["nranks"] = 2;
auto op = f::OpRegistry::CreateOp(
"c_reducescatter", {{"X", {"Data"}}}, {{"Out", {"OutData"}}}, attrs);
int iter_num = 10;
for (int i = 0; i < iter_num; i++) {
op->Run(*scope, place);
ctx.Wait();
}
std::vector<float> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
ctx.Wait();
PrintDebugInfo("output data", out_vec);
EXPECT_EQ(out_vec.size(), init.size() / 2);
for (uint32_t i = 0; i < out_vec.size(); i++) {
EXPECT_EQ(out_vec[i], 2.0);
}
}
TEST(c_reducescatter, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
// only support one device, if more than one device, use first default
p::NPUDeviceContext ctx(p::NPUPlace(atoi(FLAGS_selected_npus.c_str())));
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
TestHCCLReduceScatterOp(&scope, ctx);
}
paddle/fluid/operators/collective/c_sync_calc_stream_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP_ITSELF(elementwise_add);
USE_OP_DEVICE_KERNEL(elementwise_add, NPU);
USE_OP_DEVICE_KERNEL(c_sync_calc_stream, NPU);
template <typename T>
void Compare(f::Scope* scope, const p::DeviceContext& ctx) {
// init
auto x = scope->Var("X");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
auto y = scope->Var("Y");
auto tensor_y = y->GetMutable<phi::DenseTensor>();
std::vector<T> init_x;
for (int64_t i = 0; i < 10 * 10; ++i) {
init_x.push_back(static_cast<T>(1.0));
}
std::vector<T> init_y;
for (int64_t i = 0; i < 10 * 10; ++i) {
init_y.push_back(static_cast<T>(2.0));
}
paddle::framework::TensorFromVector(init_x, ctx, tensor_x);
tensor_x->Resize({10, 10});
paddle::framework::TensorFromVector(init_y, ctx, tensor_y);
tensor_y->Resize({10, 10});
f::AttributeMap attrs;
auto place = ctx.GetPlace();
auto out = scope->Var("Out");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
// sync data
auto sync_op0 = f::OpRegistry::CreateOp(
"c_sync_calc_stream", {{"X", {"X"}}}, {{"Out", {"Out"}}}, attrs);
sync_op0->Run(*scope, place);
// run
auto op = f::OpRegistry::CreateOp("elementwise_add",
{{"X", {"X"}}, {"Y", {"Y"}}},
{{"Out", {"Out"}}},
attrs);
op->Run(*scope, place);
// sync op run
auto sync_op = f::OpRegistry::CreateOp(
"c_sync_calc_stream", {{"X", {"X"}}}, {{"Out", {"Out"}}}, attrs);
sync_op->Run(*scope, place);
std::vector<T> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
// sync op copy
auto sync_op2 = f::OpRegistry::CreateOp(
"c_sync_calc_stream", {{"X", {"X"}}}, {{"Out", {"Out"}}}, attrs);
sync_op2->Run(*scope, place);
float expected = 3.0;
EXPECT_EQ(out_vec.size(), init_x.size());
for (uint32_t i = 0; i < out_vec.size(); i++) {
EXPECT_EQ(out_vec[i], static_cast<T>(expected));
}
}
TEST(c_sync_calc_stream, NPU_fp32) {
f::Scope scope;
p::NPUDeviceContext ctx(p::NPUPlace(0));
Compare<float>(&scope, ctx);
}
paddle/fluid/operators/collective/c_sync_comm_stream_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/c_broadcast_op.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(c_broadcast);
USE_OP_DEVICE_KERNEL(c_sync_comm_stream, NPU);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(c_broadcast, NPU);
DECLARE_string(selected_npus);
template <typename T>
void PrintDebugInfo(const std::string preStr, const std::vector<T>& data) {
std::string debugstring = "";
for (auto ele : data) {
debugstring += std::to_string(ele) + std::string(",");
}
VLOG(2) << preStr << ":" << std::endl << debugstring;
}
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
void TestHCCLBroadcastOp(f::Scope* scope, const p::DeviceContext& ctx) {
std::cout << "BEGIN TEST:" << __FUNCTION__ << std::endl;
// init
auto x = scope->Var("Data");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
int num = 2;
std::vector<float> init;
int rank_id = atoi(getenv("RANK_ID"));
std::cout << "rank_id:" << rank_id << std::endl;
for (int64_t i = 0; i < num * num; ++i) {
init.push_back(1.0 + rank_id);
std::cout << init[0];
}
std::cout << std::endl;
paddle::framework::TensorFromVector(init, ctx, tensor_x);
tensor_x->Resize({num, num});
ctx.Wait();
auto place = ctx.GetPlace();
auto out = scope->Var("OutData");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
tensor_out->Resize({num, num});
tensor_out->mutable_data<float>(place); // allocate
ctx.Wait();
// run
f::AttributeMap attrs;
attrs["tag"] = std::string("tagx");
attrs["root"] = 0;
attrs["ring_id"] = 0;
auto op = f::OpRegistry::CreateOp(
"c_broadcast", {{"X", {"Data"}}}, {{"Out", {"OutData"}}}, attrs);
op->Run(*scope, place);
// comm sync
auto sync_op = f::OpRegistry::CreateOp(
"c_sync_comm_stream", {{"X", {"Data"}}}, {{"Out", {"OutData"}}}, attrs);
sync_op->Run(*scope, place);
// ctx.Wait();
std::vector<float> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
EXPECT_EQ(out_vec.size(), init.size());
for (uint32_t i = 0; i < out_vec.size(); i++) {
EXPECT_EQ(out_vec[i], 1.0);
}
}
TEST(c_sync_comm_stream_op, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
// only support one device, if more than one device, use first default
p::NPUDeviceContext ctx(p::NPUPlace(atoi(FLAGS_selected_npus.c_str())));
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
TestHCCLBroadcastOp(&scope, ctx);
}
paddle/fluid/operators/collective/checknumeric_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <cmath>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/c_allreduce_op.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(c_allreduce_sum);
USE_OP_DEVICE_KERNEL(c_allreduce_sum, NPU);
DECLARE_string(selected_npus);
template <typename T>
bool Check(T value, int size = 2 * 512 * 8192) {
f::Scope scope;
auto x = scope.Var("in");
auto& ctx = *dynamic_cast<p::NPUDeviceContext*>(
p::DeviceContextPool::Instance().Get(p::NPUPlace(0)));
auto place = ctx.GetPlace();
auto tensor_x = x->GetMutable<phi::DenseTensor>();
tensor_x->Resize({size});
tensor_x->mutable_data<T>(place); // allocate
std::vector<T> init;
for (int64_t i = 0; i < size; ++i) {
init.push_back(static_cast<T>(value));
}
paddle::framework::TensorFromVector(init, ctx, tensor_x);
bool result = paddle::operators::ContainsNan(ctx, ctx.stream(), tensor_x);
return result;
}
TEST(check_numeric, NPU) {
auto inf = std::numeric_limits<float>::infinity();
auto fp16_inf = static_cast<p::float16>(inf);
auto nan = NAN;
auto fp16_nan = static_cast<p::float16>(nan);
bool result = false;
// Normal
VLOG(0) << "start normal";
result = Check<p::float16>(static_cast<p::float16>(65546));
ASSERT_FALSE(result);
Check<float>(static_cast<float>(1.0));
ASSERT_FALSE(result);
// Inf
VLOG(0) << "start inf";
result = Check<p::float16>(fp16_inf);
ASSERT_FALSE(result);
result = Check<float>(inf);
ASSERT_FALSE(result);
// Nan
VLOG(0) << "start nan";
result = Check<p::float16>(fp16_nan);
ASSERT_TRUE(result);
result = Check<float>(nan);
ASSERT_TRUE(result);
}
paddle/fluid/operators/collective/mp_allreduce_sum_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
// 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.
#include "paddle/fluid/operators/collective/c_allreduce_op.h"
namespace paddle {
namespace platform {
struct ASCENDPlace;
} // namespace platform
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(
mp_allreduce_sum,
ops::CAllReduceOpASCENDKernel<ops::kRedSum, int>,
ops::CAllReduceOpASCENDKernel<ops::kRedSum, int8_t>,
ops::CAllReduceOpASCENDKernel<ops::kRedSum, float>,
ops::CAllReduceOpASCENDKernel<ops::kRedSum, plat::float16>)
paddle/fluid/operators/collective/partial_allgather_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/partial_allgather_op.h"
#include "paddle/fluid/platform/collective_helper.h"
namespace paddle {
namespace operators {
template <typename T>
class CallPartialGatherOpASCENDKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"PaddlePaddle should compile with NPU."));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(partial_allgather,
ops::CallPartialGatherOpASCENDKernel<int8_t>,
ops::CallPartialGatherOpASCENDKernel<int>,
ops::CallPartialGatherOpASCENDKernel<float>,
ops::CallPartialGatherOpASCENDKernel<plat::float16>);
paddle/fluid/operators/collective/partial_recv_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/partial_recv_op.h"
#include "paddle/fluid/platform/collective_helper.h"
namespace paddle {
namespace operators {
template <typename T>
class PartialRecvOpASCENDKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"PaddlePaddle should compile with NPU."));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(partial_recv,
ops::PartialRecvOpASCENDKernel<int>,
ops::PartialRecvOpASCENDKernel<int8_t>,
ops::PartialRecvOpASCENDKernel<float>,
ops::PartialRecvOpASCENDKernel<plat::float16>);
paddle/fluid/operators/collective/partial_send_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/send_v2_op.h"
#include "paddle/fluid/platform/collective_helper.h"
namespace paddle {
namespace operators {
template <typename T>
class PartialSendOpASCENDKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"PaddlePaddle should compile with NPU."));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(partial_send,
ops::PartialSendOpASCENDKernel<int>,
ops::PartialSendOpASCENDKernel<int8_t>,
ops::PartialSendOpASCENDKernel<float>,
ops::PartialSendOpASCENDKernel<plat::float16>);
paddle/fluid/operators/collective/recv_v2_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/recv_v2_op.h"
#include "paddle/fluid/distributed/collective/process_group.h"
#include "paddle/phi/api/include/tensor.h"
namespace paddle {
namespace operators {
template <typename T>
class CRecvOpASCENDKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"PaddlePaddle should compile with NPU."));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(recv_v2,
ops::CRecvOpASCENDKernel<int>,
ops::CRecvOpASCENDKernel<int8_t>,
ops::CRecvOpASCENDKernel<float>,
ops::CRecvOpASCENDKernel<plat::float16>);
paddle/fluid/operators/collective/recv_v2_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/operators/collective/recv_v2_op.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(recv_v2);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(recv_v2, NPU);
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
void TestHcomRecvOp(f::Scope* scope, const p::DeviceContext& ctx) {
std::cout << "BEGIN TEST:" << __FUNCTION__ << std::endl;
int num = atoi(getenv("DATA_SIZE"));
EXPECT_GT(num, 0);
EXPECT_LT(num, 1 << 15);
int rank_id = atoi(getenv("RANK_ID"));
VLOG(3) << "rank_id:" << rank_id << std::endl;
ctx.Wait();
auto place = ctx.GetPlace();
auto out = scope->Var("Data");
auto tensor_out = out->GetMutable<phi::DenseTensor>();
tensor_out->Resize({num, num});
tensor_out->mutable_data<float>(place); // allocate
ctx.Wait();
f::AttributeMap attrs;
attrs["tag"] = std::string("srtest");
attrs["peer"] = atoi(getenv("SRC_RANK"));
attrs["ring_id"] = 0;
attrs["srTag"] = 0;
std::vector<int> out_shape;
out_shape.push_back(num);
out_shape.push_back(num);
attrs["out_shape"] = out_shape;
auto op = f::OpRegistry::CreateOp("recv_v2", {}, {{"Out", {"Data"}}}, attrs);
VLOG(3) << "CreateOp recv_v2";
for (int i = 0; i < 10; i++) {
op->Run(*scope, place);
}
VLOG(3) << "Run op recv_v2";
std::vector<float> out_vec;
paddle::framework::TensorToVector(*tensor_out, ctx, &out_vec);
ctx.Wait();
std::vector<float> init(num * num, 1.0 * atoi(getenv("DEST_RANK")));
EXPECT_EQ(out_vec == init, true);
}
TEST(recv_v2, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
char* npu_id = getenv("FLAGS_selected_npus");
VLOG(3) << "Select npu:" << npu_id;
p::NPUDeviceContext ctx(p::NPUPlace(atoi(npu_id)));
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
TestHcomRecvOp(&scope, ctx);
}
paddle/fluid/operators/collective/send_v2_op_npu.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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 "paddle/fluid/operators/collective/send_v2_op.h"
#include "paddle/fluid/distributed/collective/process_group.h"
#include "paddle/phi/api/include/tensor.h"
namespace paddle {
namespace operators {
template <typename T>
class CSendOpASCENDKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"PaddlePaddle should compile with NPU."));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL(send_v2,
ops::CSendOpASCENDKernel<int>,
ops::CSendOpASCENDKernel<int8_t>,
ops::CSendOpASCENDKernel<float>,
ops::CSendOpASCENDKernel<plat::float16>);
paddle/fluid/operators/collective/send_v2_op_npu_test.cc 已删除 100644 → 0
浏览文件 @ c1cad896
/* 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. */
#ifndef _WIN32
#include <unistd.h>
#endif
#include <stdio.h>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/operators/collective/gen_hccl_id_op_helper.h"
#include "paddle/fluid/operators/collective/send_v2_op.h"
#include "paddle/fluid/string/printf.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace f = paddle::framework;
namespace p = paddle::platform;
USE_OP(send_v2);
USE_NO_KERNEL_OP(c_gen_hccl_id);
USE_NO_KERNEL_OP(c_comm_init_hccl);
USE_OP_DEVICE_KERNEL(send_v2, NPU);
void PrepareUniqueId(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
std::vector<int> rank_ids{0, 1};
f::AttributeMap gen_hccl_id;
std::vector<std::string> endpointList = {"127.0.0.1:6175", "127.0.0.1:6177"};
gen_hccl_id["rank"] = rank_id;
gen_hccl_id["endpoint"] = endpointList[rank_id];
std::vector<std::string> other_endpoints = {
endpointList[rank_id == 0 ? 1 : 0]};
gen_hccl_id["other_endpoints"] = other_endpoints;
auto out = scope->Var("Out");
auto id = out->GetMutable<HcclRootInfo>();
VLOG(3) << "break";
auto comm_init_op = f::OpRegistry::CreateOp(
"c_gen_hccl_id", {}, {{"Out", {"Out"}}}, gen_hccl_id);
VLOG(3) << "break";
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
memcpy(hccl_id, id, 1024);
}
void Prepare(f::Scope* scope,
const p::DeviceContext& ctx,
HcclRootInfo* hccl_id) {
auto x = scope->Var("X");
auto id = x->GetMutable<HcclRootInfo>();
memcpy(id, hccl_id, 1024);
int rank_id = atoi(getenv("RANK_ID"));
int device_id = atoi(getenv("DEVICE_ID"));
VLOG(2) << "rank_id = " << rank_id << "; device_id = " << device_id
<< "; rank_id = " << rank_id
<< "; RANK_TABLE_FILE = " << atoi(getenv("DEVICE_ID"));
// std::vector<int> rank_ids{0, 1};
f::AttributeMap comm_init_attrs;
comm_init_attrs["ring_id"] = 0;
comm_init_attrs["rank_ids"] = 2;
comm_init_attrs["rank"] = rank_id;
comm_init_attrs["device_id"] = device_id;
// comm_init_attrs["rank_ids"] = rank_ids;
auto comm_init_op = f::OpRegistry::CreateOp(
"c_comm_init_hccl", {{"X", {"X"}}}, {}, comm_init_attrs);
auto place = ctx.GetPlace();
comm_init_op->Run(*scope, place);
ctx.Wait();
}
void TestHcomSendOp(f::Scope* scope, const p::DeviceContext& ctx) {
std::cout << "BEGIN TEST:" << __FUNCTION__ << std::endl;
auto x = scope->Var("Data");
auto tensor_x = x->GetMutable<phi::DenseTensor>();
int num = atoi(getenv("DATA_SIZE"));
EXPECT_GT(num, 0);
EXPECT_LT(num, 1 << 15);
std::vector<float> init(num * num, 1.0 * atoi(getenv("DEST_RANK")));
int rank_id = atoi(getenv("RANK_ID"));
VLOG(3) << "rank id:" << rank_id;
paddle::framework::TensorFromVector(init, ctx, tensor_x);
tensor_x->Resize({num, num});
ctx.Wait();
auto place = ctx.GetPlace();
ctx.Wait();
f::AttributeMap attrs;
attrs["tag"] = std::string("srtest");
attrs["peer"] = atoi(getenv("DEST_RANK"));
attrs["ring_id"] = 0;
attrs["srTag"] = 0;
auto op = f::OpRegistry::CreateOp("send_v2", {{"X", {"Data"}}}, {}, attrs);
for (int i = 0; i < 10; i++) {
op->Run(*scope, place);
}
VLOG(3) << "send run over";
ctx.Wait();
}
TEST(send_v2, NPU) {
f::Scope scope;
HcclRootInfo hccl_id;
char* npu_id = getenv("FLAGS_selected_npus");
VLOG(3) << "Select npu:" << npu_id;
p::NPUDeviceContext ctx(p::NPUPlace(atoi(npu_id)));
PrepareUniqueId(&scope, ctx, &hccl_id);
Prepare(&scope, ctx, &hccl_id);
TestHcomSendOp(&scope, ctx);
}
Markdown is supported
0% 或 .
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 或