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

support buffer fusion

Browse Source
pull/624/head
jjfeing 5 years ago
parent e3d5fa90de
commit d03e04b828
9 changed files with 182 additions and 66 deletions
Show Stats Download Patch File Download Diff File

10
mindspore/_extends/parallel_compile/tbe_compiler/common.py
Unescape View File

@ -122,10 +122,12 @@ def get_args(op_info, arg_type):
elif arg_type == 'attrs':
for item in op_info[arg_type]:
if 'value' not in item:
raise ValueError("Json string Errors, attr key:value not found.")
if item["name"] != "isRef":
args.append(item['value'])
if item["valid"]:
if 'value' not in item:
raise ValueError("Json string Errors, attr key:value not found.")
if item["name"] != "isRef":
args.append(item['value'])
return args

3
mindspore/ccsrc/kernel/kernel_fusion.cc
Unescape View File

@ -108,7 +108,8 @@ std::map<int32_t, KernelModPtr> KernelFusion(const std::vector<FusionScopeInfo>
}
if ((task_result != nullptr) && (strcmp(task_result, "Success") != 0)) {
MS_LOG(DEBUG) << "fuison op build failed, err log: " << task_result << " change to single op build.";
MS_LOG(INFO) << "Fusion warning: Fuison op build failed, err log: " << task_result
<< " change to single op build.";
build_failed_num++;
}
auto kernel_mod_item = build_manger->TaskFinishProcess(task_id, false);

46
mindspore/ccsrc/kernel/tbe/tbe_adapter.cc
Unescape View File

@ -153,6 +153,52 @@ void TbeAdapter::InputOrderPass(const std::string &op_name, std::vector<std::vec
}
}
void TbeAdapter::FusionInputOrderPass(const std::string &op_name, const std::vector<nlohmann::json> &inputs_list,
std::vector<nlohmann::json> *inputs_json) {
MS_EXCEPTION_IF_NULL(inputs_json);
if (input_order_adjusted_ops.find(op_name) == input_order_adjusted_ops.end()) {
(void)std::copy(inputs_list.begin(), inputs_list.end(), std::back_inserter((*inputs_json)));
} else {
if (op_name == "MinimumGrad" || op_name == "MaximumGrad") {
inputs_json->emplace_back(inputs_list[2]);
inputs_json->emplace_back(inputs_list[0]);
inputs_json->emplace_back(inputs_list[1]);
for (size_t i = 3; i < inputs_list.size(); ++i) {
inputs_json->emplace_back(inputs_list[i]);
}
} else {
inputs_json->emplace_back(inputs_list[1]);
inputs_json->emplace_back(inputs_list[0]);
for (size_t i = 2; i < inputs_list.size(); ++i) {
inputs_json->emplace_back(inputs_list[i]);
}
}
}
}
void TbeAdapter::FusionDataOrderPass(const std::string &op_name, const std::vector<AnfNodePtr> &data_layer,
std::vector<AnfNodePtr> *reorder_data_layer) {
MS_EXCEPTION_IF_NULL(reorder_data_layer);
if (input_order_adjusted_ops.find(op_name) == input_order_adjusted_ops.end()) {
(void)std::copy(data_layer.begin(), data_layer.end(), std::back_inserter((*reorder_data_layer)));
} else {
if (op_name == "MinimumGrad" || op_name == "MaximumGrad") {
reorder_data_layer->emplace_back(data_layer[2]);
reorder_data_layer->emplace_back(data_layer[0]);
reorder_data_layer->emplace_back(data_layer[1]);
for (size_t i = 3; i < data_layer.size(); ++i) {
reorder_data_layer->emplace_back(data_layer[i]);
}
} else {
reorder_data_layer->emplace_back(data_layer[1]);
reorder_data_layer->emplace_back(data_layer[0]);
for (size_t i = 2; i < data_layer.size(); ++i) {
reorder_data_layer->emplace_back(data_layer[i]);
}
}
}
}
std::map<std::string, FAttrsPass> TbeAdapter::build_json_attr_pass_map_ = {
{"MaximumGrad", TbeAdapter::MaximumGradAttrJsonPass},
{"MinimumGrad", TbeAdapter::MinimumGradAttrJsonPass},

13
mindspore/ccsrc/kernel/tbe/tbe_adapter.h
Unescape View File

@ -44,15 +44,12 @@ class TbeAdapter {
static void GenTopKV2IndicesTensorInfo(const std::shared_ptr<AnfNode> &anf_node, size_t real_input_index,
std::vector<nlohmann::json> *input_list, kCreaterType creater_type);
static void FusionInputOrderPass(const std::string &op_name, const std::vector<nlohmann::json> &inputs_list,
std::vector<nlohmann::json> *inputs_json);
static void FusionDataOrderPass(const std::string &op_name, const std::vector<AnfNodePtr> &data_layer,
std::vector<AnfNodePtr> *reorder_data_layer);
private:
static void Conv2DAttrJsonPass(const AnfNodePtr &anf_node, const std::vector<std::shared_ptr<OpAttr>> &op_info_attrs,
nlohmann::json *attrs_json);
static void Conv2DBackpropFilterAttrJsonPass(const AnfNodePtr &anf_node,
const std::vector<std::shared_ptr<OpAttr>> &op_info_attrs,
nlohmann::json *attrs_json);
static void Conv2DBackpropInputAttrJsonPass(const AnfNodePtr &anf_node,
const std::vector<std::shared_ptr<OpAttr>> &op_info_attrs,
nlohmann::json *attrs_json);
static void MaximumGradAttrJsonPass(const AnfNodePtr &anf_node,
const std::vector<std::shared_ptr<OpAttr>> &op_info_attrs,
nlohmann::json *attrs_json);

157
mindspore/ccsrc/kernel/tbe/tbe_kernel_build.cc
View File

File diff suppressed because it is too large Load Diff

15
mindspore/ccsrc/kernel/tbe/tbe_kernel_build.h
Unescape View File

@ -35,6 +35,8 @@ namespace kernel {
// kernel operate type used for generate json
class TbeKernelBuild {
enum FusionDataType { kFusionNormal = 0, kFusionAddN, kFusionReLUGradV2 };
public:
static bool GetIOSize(const nlohmann::json &kernel_json, std::vector<size_t> *input_size_list,
std::vector<size_t> *output_size_list);
@ -48,8 +50,9 @@ class TbeKernelBuild {
private:
TbeKernelBuild() = default;
~TbeKernelBuild() = default;
static bool GenFusionDataInputJson(const std::shared_ptr<mindspore::AnfNode> &data_input, nlohmann::json *data_str,
size_t *index);
static bool GenFusionDataInputJson(const std::shared_ptr<mindspore::AnfNode> &data_input,
const std::map<const AnfNodePtr, FusionDataType> &spec_data_input,
nlohmann::json *data_str, size_t *index);
static bool GenFusionComputeJson(const mindspore::AnfNodePtr &compute_node,
std::vector<std::vector<mindspore::AnfNodePtr>>::iterator *layer_iter,
nlohmann::json *compute_op_str, std::string *fusion_kernel_name, size_t *index);
@ -60,13 +63,17 @@ class TbeKernelBuild {
static bool GenFusionComputeOutputJson(const mindspore::CNodePtr &cnode,
std::vector<nlohmann::json> *output_desc_list);
static void GenDescJson(const std::shared_ptr<mindspore::AnfNode> &anf_node, size_t node_out_idx,
size_t desc_output_idx, nlohmann::json *output_desc);
size_t desc_output_idx, nlohmann::json *output_desc,
FusionDataType fusion_data_type = kFusionNormal);
static void GenReusedOutputDesc(const std::shared_ptr<mindspore::AnfNode> &anf_node, size_t index,
size_t output_index, nlohmann::json *output_desc);
static size_t GetIOSizeImpl(const nlohmann::json &desc);
static bool GetSpecInputLayers(const std::string &op_name, const std::vector<mindspore::AnfNodePtr> &reorder_layer,
std::map<const AnfNodePtr, FusionDataType> *spec_data_input);
static bool GetInputLayers(const std::vector<mindspore::AnfNodePtr> &input_nodes,
const std::vector<mindspore::AnfNodePtr> &compute_nodes,
std::vector<std::vector<mindspore::AnfNodePtr>> *input_layers);
std::vector<std::vector<mindspore::AnfNodePtr>> *input_layers,
std::map<const AnfNodePtr, FusionDataType> *spec_data_input);
static bool IsDynamicInput(const CNodePtr &cnode);
static size_t GetOptionalInput(const CNodePtr &cnode, bool is_dynamic_input);
};

2
mindspore/ccsrc/utils/utils.h
Unescape View File

@ -122,6 +122,8 @@ constexpr auto kSendOpName = "Send";
constexpr auto kRecvOpName = "Recv";
constexpr auto kReluV2OpName = "ReLUV2";
constexpr auto kReluGradV2OpName = "ReluGradV2";
constexpr auto kAddNOpName = "AddN";
constexpr auto kConv2DBackpropInputOpName = "Conv2DBackpropInput";
constexpr auto kFusionOpConv2DBackpropInputReluGradV2Name = "FusionOp_Conv2DBackpropInput_ReluGradV2";
constexpr auto kFusionOpConv2DBackpropInputAddNReluGradV2Name = "FusionOp_Conv2DBackpropInput_AddN_ReluGradV2";

1
mindspore/ops/_op_impl/tbe/reduce_mean.py
Unescape View File

@ -31,6 +31,7 @@ reduce_mean_op_info = TBERegOp("ReduceMean") \
.dtype_format(DataType.U8_Default, DataType.U8_Default) \
.dtype_format(DataType.F16_Default, DataType.F16_Default) \
.dtype_format(DataType.F32_Default, DataType.F32_Default) \
.dtype_format(DataType.F16_5HD, DataType.F16_5HD) \
.get_op_info()

1
mindspore/ops/operations/nn_ops.py
Unescape View File

@ -701,6 +701,7 @@ class Conv2D(PrimitiveWithInfer):
self.add_prim_attr('data_format', "NCHW")
self.out_channel = validator.check_integer('out_channel', out_channel, 0, Rel.GT, self.name)
self.group = validator.check_integer('group', group, 0, Rel.GT, self.name)
self.add_prim_attr('offset_a', 0)
def infer_shape(self, x_shape, w_shape):
validator.check_integer("weight rank", len(w_shape), 4, Rel.EQ, self.name)

Write Preview
Loading…
Cancel
Save