Repair nccl op test #8575
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -14,19 +14,15 @@ limitations under the License. */ | |
|
|
||
| #include <glog/logging.h> | ||
| #include <gtest/gtest.h> | ||
| #include <memory> | ||
| #include <mutex> | ||
| #include <thread> | ||
| #include <vector> | ||
|
|
||
| #include "paddle/fluid/framework/init.h" | ||
| #include "paddle/fluid/framework/op_desc.h" | ||
| #include "paddle/fluid/framework/op_registry.h" | ||
| #include "paddle/fluid/framework/program_desc.h" | ||
| #include "paddle/fluid/operators/nccl/nccl_gpu_common.h" | ||
| #include "paddle/fluid/platform/device_context.h" | ||
| #include "paddle/fluid/platform/enforce.h" | ||
|
|
@@ -41,15 +37,24 @@ USE_CUDA_ONLY_OP(ncclBcast); | |
| namespace f = paddle::framework; | ||
| namespace p = paddle::platform; | ||
|
|
||
| // test data amount | ||
| const f::DDim kDims = {20, 20}; | ||
|
|
||
| // nccl op common tester, init communicator. | ||
| class NCCLTester : public ::testing::Test { | ||
| public: | ||
| virtual void SetUp() override { | ||
| int count = p::GetCUDADeviceCount(); | ||
| if (count <= 1) { | ||
| LOG(WARNING) | ||
| << "Cannot test multi-gpu nccl, because the CUDA device count is " | ||
| << count; | ||
| exit(0); | ||
| } | ||
| for (int i = 0; i < count; ++i) { | ||
| gpu_list.emplace_back(i); | ||
| } | ||
|
|
||
| paddle::platform::CPUPlace cpu_place; | ||
| for (size_t i = 0; i < gpu_list.size(); ++i) { | ||
| p::CUDAPlace place(i); | ||
|
|
@@ -70,18 +75,24 @@ class NCCLTester : public ::testing::Test { | |
| std::unique_ptr<f::OpDesc> op1(new f::OpDesc); | ||
|
|
||
| op1->SetType("ncclInit"); | ||
| op1->SetInput("parallel_scopes", {"p_scopes"}); | ||
| op1->SetOutput("Communicator", {"comm"}); | ||
|
|
||
| auto *var = g_scope.Var("comm"); | ||
| var->GetMutable<p::Communicator>(); | ||
|
|
||
| auto *scope_var = g_scope.Var("p_scopes"); | ||
| auto *p_scopes = scope_var->GetMutable<std::vector<f::Scope *>>(); | ||
| (*p_scopes).resize(gpu_list.size()); | ||
|
|
||
| auto op = f::OpRegistry::CreateOp(*op1); | ||
| VLOG(1) << "invoke NCCLInitOp."; | ||
| op->Run(g_scope, cpu_place); | ||
| VLOG(1) << "NCCLInitOp finished."; | ||
| } | ||
|
|
||
| int GetGPUData(int gpu_id) { return gpu_id + 42; } | ||
|
|
||
| template <class T> | ||
| void PerThreadProgram(int gpu_id, const f::OpDesc &op_desc, f::Scope *scope) { | ||
| std::unique_lock<std::mutex> lk(mu); | ||
|
|
@@ -97,7 +108,7 @@ class NCCLTester : public ::testing::Test { | |
| send_tensor->Resize(kDims); | ||
| send_tensor->mutable_data<T>(kDims, place); | ||
|
There was a problem hiding this comment. Maybe line 108 is unnecessary. |
||
|
|
||
| std::vector<T> send_vector(f::product(kDims), GetGPUData(gpu_id)); | ||
| paddle::framework::TensorFromVector<T>(send_vector, *ctx, send_tensor); | ||
| ctx->Wait(); | ||
|
There was a problem hiding this comment. Is it necessary to synchronize here? I think the copying will synchronize the GPU and CPU in line 179. There was a problem hiding this comment. The copying is Paddle/paddle/fluid/memory/memcpy.cc Lines 30 to 36 in 0d49b92 There was a problem hiding this comment. I don't think so because the memory is pageable in CPU side, the copying doesn't return immediately until the copy has completed. There was a problem hiding this comment. |
||
| VLOG(1) << "Send Tensor filled with elements " << send_tensor->numel(); | ||
|
|
@@ -121,27 +132,11 @@ class NCCLTester : public ::testing::Test { | |
| std::vector<p::DeviceContext *> dev_ctxs; | ||
| f::Scope g_scope; | ||
| std::mutex mu; | ||
| std::vector<int> gpu_list; | ||
|
There was a problem hiding this comment. Data members of classes should be with a trailing underscore. |
||
| }; | ||
|
|
||
| // ncclInitOp with desc | ||
| TEST_F(NCCLTester, ncclInitOp) {} | ||
|
|
||
| // ncclAllReduceOp with desc | ||
| TEST_F(NCCLTester, ncclAllReduceOp) { | ||
|
|
@@ -166,8 +161,10 @@ TEST_F(NCCLTester, ncclAllReduceOp) { | |
| ths[i].join(); | ||
| } | ||
|
|
||
| float expected_result = 0.0; | ||
| for (int gpu_id : gpu_list) { | ||
| expected_result = expected_result + GetGPUData(gpu_id); | ||
| } | ||
|
|
||
| for (size_t i = 0; i < dev_scopes.size(); ++i) { | ||
| p::CPUPlace cpu_place; | ||
|
|
@@ -185,7 +182,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) { | |
| static_cast<p::CUDADeviceContext *>(dev_ctxs[i])->stream()); | ||
|
|
||
| for (int64_t j = 0; j < f::product(kDims); ++j) { | ||
| ASSERT_NEAR(ct[j], expected_result, 1e-5); | ||
| } | ||
| } | ||
| } | ||
|
|
@@ -215,8 +212,10 @@ TEST_F(NCCLTester, ncclReduceOp) { | |
| ths[i].join(); | ||
| } | ||
|
|
||
| float expected_result = 0.0; | ||
| for (int gpu_id : gpu_list) { | ||
| expected_result = expected_result + GetGPUData(gpu_id); | ||
| } | ||
|
|
||
| p::CPUPlace cpu_place; | ||
| p::CUDAPlace gpu_place(gpu_list[kRoot]); | ||
|
|
@@ -234,7 +233,7 @@ TEST_F(NCCLTester, ncclReduceOp) { | |
| static_cast<p::CUDADeviceContext *>(dev_ctxs[kRoot])->stream()); | ||
|
|
||
| for (int64_t j = 0; j < f::product(kDims); ++j) { | ||
| ASSERT_NEAR(ct[j], expected_result, 1e-5); | ||
| } | ||
| } | ||
|
|
||
|
|
@@ -264,8 +263,7 @@ TEST_F(NCCLTester, ncclBcastOp) { | |
| } | ||
|
|
||
| const int idx = 1; | ||
| float result = GetGPUData(kRoot); | ||
|
|
||
| p::CPUPlace cpu_place; | ||
| p::CUDAPlace gpu_place(gpu_list[idx]); | ||
|
|
@@ -285,34 +283,3 @@ TEST_F(NCCLTester, ncclBcastOp) { | |
| ASSERT_NEAR(ct[j], result, 1e-5); | ||
| } | ||
| } | ||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This function is necessary, because
Simply set GPU data = gpu_id won't expose the incorrect linking error. More specifically, Paddle might be compiled with NCCL1.3 header while dynamically linked with NCCL2.so.