Skip to content

Commit

Permalink
Merge remote-tracking branch 'upstream/develop' into faster-roi-clean
Browse files Browse the repository at this point in the history
  • Loading branch information
@tsocha
tsocha committed Dec 15, 2021
2 parents 3f898c0 + 5082249 commit 0ab9a06
Show file tree
Hide file tree
Showing 65 changed files with 1,455 additions and 479 deletions.
2 changes: 1 addition & 1 deletion cmake/third_party.cmake
View file
Original file line number Diff line number Diff line change
Expand Up @@ -393,7 +393,7 @@ endif (WIN32)

if (WITH_INFRT)
include(external/llvm)
list(APPEND third_party_deps external_llvm)
list(APPEND third_party_deps ${llvm_libs})
endif()

if (WITH_IPU)
Expand Down
37 changes: 11 additions & 26 deletions paddle/fluid/eager/auto_code_generator/eager_generator.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1174,7 +1174,7 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
FWD_INS_CONTENT_TEMPLATE, input_name, input_name, input_name);
} else {
const char* FWD_INS_CONTENT_TEMPLATE =
" if(%s.initialized()) "
" if(%s.safe_initialized()) "
"ins[\"%s\"] = egr::EagerUtils::SyncToVars(%s)\n;";
generated_function_body += paddle::string::Sprintf(
FWD_INS_CONTENT_TEMPLATE, input_name, input_name, input_name);
Expand All @@ -1196,25 +1196,21 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
// in form of shared_ptr<EagerTensor>/vector<shared_ptr<EagerTensor>>
if (output.duplicable()) {
const char* FWD_NUM_ARG_TEMPLATE =
", std::vector<std::shared_ptr<egr::EagerTensor>>& %s";
", std::vector<egr::EagerTensor>& %s";
std::string arg_str =
paddle::string::Sprintf(FWD_NUM_ARG_TEMPLATE, output_var_name);
dygraph_function_args_str += arg_str;

const char* FWD_OUTS_CONTENT_TEMPLATE = "{ \"%s\", %s },";
outs_contents_str += paddle::string::Sprintf(
FWD_OUTS_CONTENT_TEMPLATE, output_name, output_var_name);
} else {
const char* FWD_NUM_ARG_TEMPLATE =
", std::shared_ptr<egr::EagerTensor>& %s";
const char* FWD_NUM_ARG_TEMPLATE = ", egr::EagerTensor& %s";
std::string arg_str =
paddle::string::Sprintf(FWD_NUM_ARG_TEMPLATE, output_var_name);
dygraph_function_args_str += arg_str;

const char* FWD_OUTS_CONTENT_TEMPLATE = "{ \"%s\", {%s} },";
outs_contents_str += paddle::string::Sprintf(
FWD_OUTS_CONTENT_TEMPLATE, output_name, output_var_name);
}
const char* FWD_OUTS_CONTENT_TEMPLATE =
"{ \"%s\", egr::EagerUtils::TrySyncToVars(&%s) },";
outs_contents_str += paddle::string::Sprintf(
FWD_OUTS_CONTENT_TEMPLATE, output_name, output_var_name);

} else {
if (output.duplicable()) {
Expand Down Expand Up @@ -1557,22 +1553,11 @@ static std::string GenerateGradNodeCCContents(
"fwd_outputs_name_pos_map"));

size_t grads_position = fwd_outputs_name_pos_map.at(fwd_name);
std::string grad_ptr_name = fwd_name + "_ptrs";
const char* GET_GRADS_PTR_TEMPLATE =
" std::vector<std::shared_ptr<egr::EagerTensor>> %s;\n"
" for(const auto& t : grads[%d]) {\n "
"%s.emplace_back(std::move(std::make_shared<egr::EagerTensor>(t))"
");"
"\n }\n";
std::string grads_ptr_str =
paddle::string::Sprintf(GET_GRADS_PTR_TEMPLATE, grad_ptr_name,
grads_position, grad_ptr_name);
generated_grad_function_body += grads_ptr_str;
generated_grad_function_body += "\n";

const char* GRAD_OUTS_CONTENT_TEMPLATE = "{ \"%s\", %s },";

const char* GRAD_OUTS_CONTENT_TEMPLATE =
"{ \"%s\", egr::EagerUtils::SyncToVars(grads[%d]) },";
outs_contents_str += paddle::string::Sprintf(
GRAD_OUTS_CONTENT_TEMPLATE, grad_output_name, grad_ptr_name);
GRAD_OUTS_CONTENT_TEMPLATE, grad_output_name, grads_position);

} else {
size_t fwd_input_position = fwd_inputs_name_pos_map.at(fwd_name);
Expand Down
4 changes: 4 additions & 0 deletions paddle/fluid/eager/eager_tensor.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -152,6 +152,10 @@ class EagerTensor final {
*/
bool initialized() const { return tensor_->initialized(); }

bool safe_initialized() const {
return initialized() || var_.IsInitialized();
}

/**
* @description: Reset the Tensor implementation
* @param None
Expand Down
24 changes: 24 additions & 0 deletions paddle/fluid/eager/utils.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -135,6 +135,30 @@ std::vector<std::shared_ptr<egr::EagerTensor>> EagerUtils::SyncToVars(
return res;
}

static std::shared_ptr<egr::EagerTensor> TrySyncToVar(
egr::EagerTensor* tensor) {
if (tensor->initialized() || tensor->Var().IsInitialized()) {
tensor->SyncToVar(paddle::framework::proto::VarType_Type_LOD_TENSOR);
}
return std::make_shared<EagerTensor>(*tensor);
}

std::vector<std::shared_ptr<egr::EagerTensor>> EagerUtils::TrySyncToVars(
egr::EagerTensor* tensor) {
return {TrySyncToVar(tensor)};
}

std::vector<std::shared_ptr<egr::EagerTensor>> EagerUtils::TrySyncToVars(
std::vector<egr::EagerTensor>* tensors) {
std::vector<std::shared_ptr<EagerTensor>> res;
size_t num = tensors->size();
res.reserve(num);
for (size_t i = 0; i < num; i++) {
res.emplace_back(TrySyncToVar(&(*tensors)[i]));
}
return res;
}

/* ---- VarBase -> Tensor ---- */
std::vector<std::shared_ptr<egr::EagerTensor>> EagerUtils::SyncToTensors(
const egr::EagerTensor& tensor) {
Expand Down
5 changes: 5 additions & 0 deletions paddle/fluid/eager/utils.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -145,6 +145,11 @@ class EagerUtils {
const std::shared_ptr<GradNodeBase>& grad_node);

// Intermidate needed remove this once we don't need legacy
static std::vector<std::shared_ptr<egr::EagerTensor>> TrySyncToVars(
egr::EagerTensor* tensor);
static std::vector<std::shared_ptr<egr::EagerTensor>> TrySyncToVars(
std::vector<egr::EagerTensor>* tensors);

static std::vector<std::shared_ptr<egr::EagerTensor>> SyncToVars(
const egr::EagerTensor& tensor);
static std::vector<std::shared_ptr<egr::EagerTensor>> SyncToVars(
Expand Down
14 changes: 7 additions & 7 deletions paddle/fluid/framework/custom_operator.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -110,8 +110,8 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx,
const std::vector<std::string>& outputs,
const std::vector<std::string>& attrs) {
VLOG(1) << "Custom Operator: Start run KernelFunc.";
std::vector<paddle::Tensor> custom_ins;
std::vector<std::vector<paddle::Tensor>> custom_vec_ins;
std::vector<paddle::experimental::Tensor> custom_ins;
std::vector<std::vector<paddle::experimental::Tensor>> custom_vec_ins;
for (auto& in_name : inputs) {
VLOG(1) << "Custom Operator: input name - " << in_name;
if (detail::IsDuplicableVar(in_name)) {
Expand All @@ -120,7 +120,7 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx,
PADDLE_ENFORCE_NE(vec_x.empty(), true,
platform::errors::NotFound(
"Input vector<tensor> (%s) is empty.", in_name));
std::vector<paddle::Tensor> custom_vec_in;
std::vector<paddle::experimental::Tensor> custom_vec_in;
for (size_t i = 0; i < vec_x.size(); ++i) {
auto* x = vec_x[i];
PADDLE_ENFORCE_NOT_NULL(
Expand All @@ -132,7 +132,7 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx,
"The %d-th tensor in input vector<tensor> (%s) "
"is not initialized.",
i, in_name));
paddle::Tensor custom_t;
paddle::experimental::Tensor custom_t;
custom_t.set_impl(std::move(experimental::MakePtenDenseTensor(*x)));
custom_vec_in.emplace_back(custom_t);
}
Expand All @@ -144,7 +144,7 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx,
PADDLE_ENFORCE_EQ(x->IsInitialized(), true,
platform::errors::InvalidArgument(
"Input tensor (%s) is not initialized.", in_name));
paddle::Tensor custom_in;
paddle::experimental::Tensor custom_in;
custom_in.set_impl(std::move(experimental::MakePtenDenseTensor(*x)));
custom_ins.emplace_back(custom_in);
}
Expand Down Expand Up @@ -207,14 +207,14 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx,
"Tensors.",
vec_true_outs.size(), outs.size()));
for (size_t j = 0; j < vec_true_outs.size(); ++j) {
experimental::MovesStorage(
experimental::MovesSharedStorage(
std::dynamic_pointer_cast<pten::DenseTensor>(outs.at(j).impl())
.get(),
vec_true_outs.at(j));
}
} else {
auto* true_out = ctx.Output<Tensor>(out_name);
experimental::MovesStorage(
experimental::MovesSharedStorage(
std::dynamic_pointer_cast<pten::DenseTensor>(outs.at(i).impl())
.get(),
true_out);
Expand Down
17 changes: 17 additions & 0 deletions paddle/fluid/framework/ir/graph_pattern_detector.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2412,6 +2412,23 @@ PDNode *patterns::OrphanedBfloat16::operator()() {
return next_op;
}

PDNode *patterns::UnsupportedBfloat16::operator()() {
auto *prev_op = pattern->NewNode(prev_op_repr())->assert_is_op();
prev_op->assert_more([&](Node *node) {
return node->Op()->HasAttr("mkldnn_data_type") == false;
});
auto *prev_out = pattern->NewNode(prev_out_repr())->AsOutput();

auto *op = pattern->NewNode(op_repr())->assert_is_op();
op->assert_more([&](Node *node) {
return node->Op()->GetAttrIfExists<std::string>("mkldnn_data_type") ==
"bfloat16";
});
prev_op->LinksTo({prev_out});
op->LinksFrom({prev_out});
return op;
}

PDNode *patterns::LastBfloat16Ops::operator()() {
auto *op = pattern->NewNode(op_repr())->assert_is_op();
op->assert_more([&](Node *node) {
Expand Down
10 changes: 10 additions & 0 deletions paddle/fluid/framework/ir/graph_pattern_detector.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1416,6 +1416,16 @@ struct OrphanedBfloat16 : public PatternBase {
PATTERN_DECL_NODE(next_op);
};

struct UnsupportedBfloat16 : public PatternBase {
UnsupportedBfloat16(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "unsupported_bfloat16") {}
PDNode* operator()();

PATTERN_DECL_NODE(prev_op);
PATTERN_DECL_NODE(prev_out);
PATTERN_DECL_NODE(op);
};

struct LastBfloat16Ops : public PatternBase {
LastBfloat16Ops(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "last_bfloat16_ops") {}
Expand Down
4 changes: 1 addition & 3 deletions paddle/fluid/framework/ir/ipu/infer_shape_pass.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,14 +13,12 @@
// limitations under the License.

#include "paddle/fluid/framework/ir/ipu/infer_shape_pass.h"

#include "paddle/fluid/platform/device/ipu/ipu_backend.h"

#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/ir/pass_tester_helper.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/platform/device/ipu/ipu_backend.h"

namespace paddle {
namespace framework {
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/framework/ir/mkldnn/conv_concat_relu_mkldnn_fuse_pass.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ ConvConcatReLUFusePass::ConvConcatReLUFusePass() {
.IsType<std::vector<int>>()
.End()
.AddAttr("data_format")
.IsStringIn({"NCHW", "NHWC", "AnyLayout"})
.IsStringIn({"NCHW"})
.End();

AddOpCompat(OpCompat("concat"))
Expand Down
21 changes: 21 additions & 0 deletions paddle/fluid/framework/ir/mkldnn/cpu_bfloat16_placement_pass.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -71,10 +71,31 @@ void CPUBfloat16PlacementPass::RemoveOrphanedOperators(
gpd(graph, handler);
}

void CPUBfloat16PlacementPass::RemoveUnsupportedOperators(
ir::Graph* graph, int* bfloat16_operators) const {
// now quantize is supported FP32 only, so try to find
// bfloat16 operator that input type is not FP32
GraphPatternDetector gpd;
patterns::UnsupportedBfloat16 unsupported_bfloat16_pattern{
gpd.mutable_pattern(), "unsupported_bfloat16"};
unsupported_bfloat16_pattern();
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_IR_NODE_FROM_SUBGRAPH(prev_out, prev_out, unsupported_bfloat16_pattern);
GET_IR_NODE_FROM_SUBGRAPH(op, op, unsupported_bfloat16_pattern);
if ((prev_out->Var()->GetDataType() != proto::VarType::FP32)) {
op->Op()->SetAttr("mkldnn_data_type", std::string("float32"));
bfloat16_operators--;
}
};
gpd(graph, handler);
}

void CPUBfloat16PlacementPass::ApplyImpl(ir::Graph* graph) const {
int bfloat16_operators = 0;
SetMkldnnDataType(graph, &bfloat16_operators);
RemoveOrphanedOperators(graph, &bfloat16_operators);
RemoveUnsupportedOperators(graph, &bfloat16_operators);
PrettyLogDetail("--- marked %d operators to bfloat16 ",
bfloat16_operators);
}
Expand Down
3 changes: 3 additions & 0 deletions paddle/fluid/framework/ir/mkldnn/cpu_bfloat16_placement_pass.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,9 @@ class CPUBfloat16PlacementPass : public Pass {

void RemoveOrphanedOperators(ir::Graph* graph, int* bfloat16_operators) const;

void RemoveUnsupportedOperators(ir::Graph* graph,
int* bfloat16_operators) const;

void ApplyImpl(ir::Graph* graph) const override;
};

Expand Down
54 changes: 53 additions & 1 deletion paddle/fluid/framework/new_executor/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -12,4 +12,56 @@ cc_library(stream_analyzer SRCS stream_analyzer.cc DEPS ${DEVICE_EVENT_LIBS} glo
cc_library(interpretercore SRCS interpretercore.cc DEPS workqueue ${DEVICE_EVENT_LIBS} interpretercore_util interpretercore_garbage_collector stream_analyzer event_manager)
cc_library(standalone_executor SRCS standalone_executor.cc DEPS interpretercore)
cc_test(workqueue_test SRCS workqueue_test.cc DEPS workqueue)
# cc_binary(standalone_executor_test SRCS standalone_executor_test.cc DEPS interpretercore standalone_executor operator op_registry executor ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} profiler)

# skip win32 since wget is not installed by default on windows machine.
# skip COVERAGE_CI since the test runs slowly because of instrumentation.
if (WITH_TESTING AND NOT WIN32 AND NOT WITH_COVERAGE)
add_custom_target(
download_program
COMMAND wget -nc https://paddle-ci.gz.bcebos.com/new_exec/lm_main_program
COMMAND wget -nc https://paddle-ci.gz.bcebos.com/new_exec/lm_startup_program
)

# all operators used in the program
set(OPS
fill_constant_op
uniform_random_op
lookup_table_op
transpose_op
reshape_op
split_op
slice_op
concat_op
matmul_op
elementwise_add_op
elementwise_mul_op
softmax_with_cross_entropy_op
reduce_mean_op
reduce_sum_op
activation_op
sum_op
elementwise_max_op
elementwise_div_op
sgd_op
squared_l2_norm_op
memcpy_h2d_op
memcpy_d2h_op)

# All deps of the operators above, part of GLOB_OPERATOR_DEPS.
set(OP_DEPS
generator
softmax
selected_rows_functor
jit_kernel_helper
concat_and_split
cross_entropy)

cc_test(standalone_executor_test SRCS standalone_executor_test.cc DEPS interpretercore standalone_executor operator op_registry executor ${OPS} ${OP_DEPS})
set_tests_properties(standalone_executor_test PROPERTIES TIMEOUT 100)

add_dependencies(standalone_executor_test download_program)
if (WITH_PROFILER)
target_link_libraries(standalone_executor_test profiler)
add_dependencies(standalone_executor_test profiler)
endif()
endif()
Loading

1 comment on commit 0ab9a06

@paddle-bot-old
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅Congratulation! Your pull request passed all required CI. You could ask reviewer(s) to approve and merge. 🎉

Please sign in to comment.