m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix cpu_bfloat16_pass (#28730)

Browse Source 
* Fix cpu_bfloat16_pass

* Add output_format

* Fix incorrect SetOutput

* Change fromating
musl/disable_test_yolov3_temporarily
joanna.wozna.intel 5 years ago committed by GitHub
parent 2fd16cf6fc
commit fddea67445
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 315 additions and 109 deletions
Show Stats Download Patch File Download Diff File

30
paddle/fluid/framework/ir/graph_pattern_detector.cc
Unescape View File

@ -2181,6 +2181,36 @@ PDNode *patterns::FirstBfloat16Ops::operator()() {
return op;
}
PDNode *patterns::DuplicatedInputs::operator()() {
auto op = pattern->NewNode(op_repr())->assert_is_ops({"concat", "sum"});
op->assert_more([&](Node *node) {
return node->Op()->GetAttrIfExists<std::string>("mkldnn_data_type") ==
"bfloat16";
});
return op;
}
PDNode *patterns::UnnecessaryReorders::operator()() {
auto prev_op = pattern->NewNode(prev_op_repr())->assert_is_op();
prev_op->assert_more([&](Node *node) {
return node->Op()->GetAttrIfExists<std::string>("mkldnn_data_type") ==
"bfloat16";
});
auto *quant_in = pattern->NewNode(quant_in_repr())
->assert_is_op_input("quantize", "Input");
auto *quant_op = pattern->NewNode(quant_op_repr())->assert_is_op("quantize");
auto *quant_out = pattern->NewNode(quant_out_repr())
->assert_is_op_output("quantize", "Output");
prev_op->LinksTo({quant_in});
quant_op->LinksFrom({quant_in}).LinksTo({quant_out});
return quant_out;
}
PDNode *patterns::MKLDNNInPlace::operator()() {
const std::unordered_set<std::string> &supported_op_types = {
"abs",

20
paddle/fluid/framework/ir/graph_pattern_detector.h
Unescape View File

@ -1273,6 +1273,26 @@ struct FirstBfloat16Ops : public PatternBase {
PATTERN_DECL_NODE(op);
};
struct DuplicatedInputs : public PatternBase {
DuplicatedInputs(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "many_inputs_op") {}
PDNode* operator()();
PATTERN_DECL_NODE(op);
};
struct UnnecessaryReorders : public PatternBase {
UnnecessaryReorders(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "unnecessary_reorders") {}
PDNode* operator()();
PATTERN_DECL_NODE(prev_op);
PATTERN_DECL_NODE(quant_in);
PATTERN_DECL_NODE(quant_op);
PATTERN_DECL_NODE(quant_out);
};
// Pattern used for enforcing inplace computation for in-place computation
// supporting DNNL ops. softmax, batch_norm and layer_norm
struct MKLDNNInPlace : public PatternBase {

218
paddle/fluid/framework/ir/mkldnn/cpu_bfloat16_pass.cc
View File

File diff suppressed because it is too large Load Diff

156
paddle/fluid/framework/ir/mkldnn/cpu_bfloat16_pass_tester.cc
Unescape View File

@ -42,60 +42,45 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
type == "dropout") {
op->SetInput("X", {inputs[0]});
op->SetOutput("Out", {outputs[0]});
op->SetAttr("mkldnn_data_type", mkldnn_data_type);
if (type != "dropout") op->SetAttr("mkldnn_data_type", mkldnn_data_type);
} else if (type == "fc") {
op->SetInput("Input", {inputs[0]});
op->SetOutput("Out", {outputs[0]});
op->SetAttr("mkldnn_data_type", mkldnn_data_type);
} else if (type == "concat") {
op->SetAttr("force_fp32_output", force_fp32_output);
} else if (type == "concat" || type == "sum") {
op->SetInput("X", inputs);
op->SetOutput("Out", outputs);
op->SetAttr("mkldnn_data_type", mkldnn_data_type);
} else if (type == "matmul" || type == "elementwise_add") {
} else if (type == "matmul" || type == "elementwise_add" ||
type == "elementwise_mul") {
op->SetInput("X", {inputs[0]});
if (inputs.size() > 1) op->SetInput("Y", {inputs[1]});
op->SetOutput("Out", {outputs[0]});
op->SetAttr("mkldnn_data_type", mkldnn_data_type);
if (type == "matmul") op->SetAttr("force_fp32_output", force_fp32_output);
} else if (type == "layer_norm") {
op->SetInput("X", {inputs[0]});
op->SetOutput("Y", {outputs[0]});
op->SetAttr("mkldnn_data_type", mkldnn_data_type);
}
}
static const std::initializer_list<std::string> variable_names{
"z", "a", "b", "c", "d", "e", "f", "g", "h", "i"};
void PreparePass(std::unique_ptr<ir::Graph>* graph, const ProgramDesc& prog,
const std::initializer_list<std::string> variable_names,
int* original_nodes_num, int* current_nodes_num) {
auto pass = PassRegistry::Instance().Get("cpu_bfloat16_pass");
graph->reset(pass->Apply(graph->release()));
*original_nodes_num = (*graph)->Nodes().size();
(*graph).reset(pass->Apply((*graph).release()));
*current_nodes_num = (*graph)->Nodes().size();
}
static const std::initializer_list<std::string> variable_names{
"z", "a", "b", "c", "d", "e", "f", "g", "h", "i"};
ProgramDesc BuildProgramDesc(bool use_mkldnn) {
ProgramDesc prog;
for (auto& v : variable_names) {
prog.MutableBlock(0)->Var(v);
}
SetOp(&prog, "dropout", "Dropout1", {"z"}, {"a"}, use_mkldnn, "float32");
SetOp(&prog, "conv2d", "Conv1", {"a"}, {"b"}, use_mkldnn, "bfloat16");
SetOp(&prog, "pool2d", "Pool1", {"b"}, {"c"}, use_mkldnn, "bfloat16");
SetOp(&prog, "conv2d", "Conv1", {"c"}, {"d"}, use_mkldnn, "bfloat16");
SetOp(&prog, "dropout", "Dropout2", {"d"}, {"e"}, use_mkldnn, "float32");
SetOp(&prog, "transpose2", "Transpose1", {"e"}, {"f"}, use_mkldnn,
"bfloat16");
SetOp(&prog, "reshape2", "Reshape1", {"f"}, {"g"}, use_mkldnn, "bfloat16");
SetOp(&prog, "concat", "Concat1", {"g"}, {"h"}, use_mkldnn, "bfloat16");
SetOp(&prog, "dropout", "Dropout3", {"h"}, {"i"}, use_mkldnn, "float32");
return prog;
}
void MainTest(const ProgramDesc& prog, int conv_count, int pool_count,
int transpose_count, int quant_count, int dequant_count,
int added_nodes_count) {
void MainTest(const ProgramDesc& prog, int quant_count, int dequant_count,
int force_fp32_count, int added_nodes_count) {
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
int original_nodes_num, current_nodes_num;
PreparePass(&graph, prog, variable_names, &original_nodes_num,
@ -103,39 +88,114 @@ void MainTest(const ProgramDesc& prog, int conv_count, int pool_count,
int quantize_nodes_count = 0;
int dequantize_nodes_count = 0;
int conv2d_nodes_count = 0;
int pool2d_nodes_count = 0;
int transpose2_nodes_count = 0;
int force_fp32_nodes_count = 0;
for (auto* node : graph->Nodes()) {
if (node->IsOp()) {
auto* op = node->Op();
if (op->Type() == "conv2d") {
conv2d_nodes_count++;
} else if (op->Type() == "pool2d") {
pool2d_nodes_count++;
} else if (op->Type() == "transpose2") {
transpose2_nodes_count++;
} else if (op->Type() == "quantize") {
if (op->Type() == "quantize") {
quantize_nodes_count++;
} else if (op->Type() == "dequantize") {
dequantize_nodes_count++;
} else if (op->Type() == "conv2d" || op->Type() == "matmul" ||
op->Type() == "fc") {
if (op->GetAttrIfExists<bool>("force_fp32_output"))
force_fp32_nodes_count++;
}
}
}
EXPECT_EQ(conv2d_nodes_count, conv_count);
EXPECT_EQ(pool2d_nodes_count, pool_count);
EXPECT_EQ(transpose2_nodes_count, transpose_count);
EXPECT_EQ(quantize_nodes_count, quant_count);
EXPECT_EQ(dequantize_nodes_count, dequant_count);
EXPECT_EQ(force_fp32_nodes_count, force_fp32_count);
EXPECT_EQ(original_nodes_num + added_nodes_count, current_nodes_num);
}
TEST(CpuQuantizePass, quantize) {
ProgramDesc BuildProgramDescConv(bool use_mkldnn) {
ProgramDesc prog;
for (auto& v : variable_names) {
prog.MutableBlock(0)->Var(v);
}
SetOp(&prog, "dropout", "Dropout", {"a"}, {"b"}, use_mkldnn, "float32");
SetOp(&prog, "conv2d", "Conv1", {"b"}, {"c"}, use_mkldnn, "bfloat16");
SetOp(&prog, "pool2d", "Pool", {"c"}, {"d"}, use_mkldnn, "bfloat16");
SetOp(&prog, "conv2d", "Conv2", {"d"}, {"e"}, use_mkldnn, "bfloat16");
SetOp(&prog, "transpose2", "Transpose", {"e"}, {"f"}, use_mkldnn, "float32");
return prog;
}
TEST(CpuBfloat16Pass, convolution) {
bool use_mkldnn = true;
// 0 added + 1 force_fp32_output
int added_nodes = 0;
MainTest(BuildProgramDescConv(use_mkldnn), 0, 0, 1, added_nodes);
}
ProgramDesc BuildProgramDescDoubleInput(bool use_mkldnn) {
ProgramDesc prog;
for (auto& v : variable_names) {
prog.MutableBlock(0)->Var(v);
}
SetOp(&prog, "dropout", "Dropout", {"a"}, {"b"}, use_mkldnn, "float32");
SetOp(&prog, "matmul", "Matmul", {"b", "b"}, {"c"}, use_mkldnn, "bfloat16");
SetOp(&prog, "transpose2", "Transpose", {"d"}, {"e"}, use_mkldnn, "float32");
SetOp(&prog, "elementwise_add", "ElemetwiseAdd", {"c", "e"}, {"f"},
use_mkldnn, "bfloat16");
SetOp(&prog, "reshape2", "Reshape", {"f"}, {"g"}, use_mkldnn, "bfloat16");
return prog;
}
TEST(CpuBfloat16Pass, double_input_ops) {
bool use_mkldnn = true;
// 2 quant + 2 quant out
int added_nodes = 4;
MainTest(BuildProgramDescDoubleInput(use_mkldnn), 2, 0, 0, added_nodes);
}
ProgramDesc BuildProgramDescDuplicatedInput(bool use_mkldnn) {
ProgramDesc prog;
for (auto& v : variable_names) {
prog.MutableBlock(0)->Var(v);
}
SetOp(&prog, "dropout", "Dropout1", {"a"}, {"b"}, use_mkldnn, "float32");
SetOp(&prog, "dropout", "Dropout2", {"c"}, {"d"}, use_mkldnn, "float32");
SetOp(&prog, "concat", "Concat", {"b", "d"}, {"e"}, use_mkldnn, "bfloat16");
SetOp(&prog, "transpose2", "Transpose", {"f"}, {"g"}, use_mkldnn, "float32");
SetOp(&prog, "sum", "Sum", {"e", "g"}, {"h"}, use_mkldnn, "bfloat16");
SetOp(&prog, "reshape2", "Reshape", {"h"}, {"i"}, use_mkldnn, "bfloat16");
return prog;
}
TEST(CpuBfloat16Pass, duplicated_input_ops) {
bool use_mkldnn = true;
// 3 quant + 3 quant out
int added_nodes = 6;
MainTest(BuildProgramDescDuplicatedInput(use_mkldnn), 3, 0, 0, added_nodes);
}
ProgramDesc BuildProgramDescDoubleOutputs(bool use_mkldnn) {
ProgramDesc prog;
for (auto& v : variable_names) {
prog.MutableBlock(0)->Var(v);
}
SetOp(&prog, "layer_norm", "LayerNorm1", {"a"}, {"b"}, use_mkldnn,
"bfloat16");
SetOp(&prog, "dropout", "Dropout1", {"b"}, {"c"}, use_mkldnn, "float32");
SetOp(&prog, "transpose2", "Transpose", {"b"}, {"d"}, use_mkldnn, "bfloat16");
SetOp(&prog, "layer_norm", "LayerNorm2", {"d"}, {"e"}, use_mkldnn,
"bfloat16");
SetOp(&prog, "reshape2", "Reshape", {"e"}, {"f"}, use_mkldnn, "float32");
SetOp(&prog, "dropout", "Dropout2", {"e"}, {"g"}, use_mkldnn, "float32");
return prog;
}
TEST(CpuBfloat16Pass, double_outputs_ops) {
bool use_mkldnn = true;
// 1 quantize + 1 dequantize
int added_nodes = 2;
MainTest(BuildProgramDesc(use_mkldnn), 2, 1, 1, 1, 2, added_nodes);
// 3 dequant + 3 dequant out
int added_nodes = 6;
MainTest(BuildProgramDescDoubleOutputs(use_mkldnn), 0, 3, 0, added_nodes);
}
} // namespace ir

Write Preview
Loading…
Cancel
Save