Star1332
2075
代码
代码
Issues785
Pull Requests671
Actions
项目讨论
Wiki
分析
Star1332
2075
已合并
perf(layer_norm): check fusion pass platform before pattern construction #11299
rk创建于 9月28日
perf(layer_norm): check fusion pass platform before pattern construction #11299
已合并
rk创建于 9月28日
从wangpengbo26/ops-nnln_onnx33合入到cann/ops-nnmaster
共 6 个文件变更+18-18
Mnorm/layer_norm/op_graph/fusion_pass/layer_norm_fusion_pass.cpp+3-3
@@ -309,6 +309,9 @@ PatternUniqPtr MakePattern(size_t config, bool form_b)
309 309 
310std::vector<PatternUniqPtr> LayerNormFusionPass::Patterns()310std::vector<PatternUniqPtr> LayerNormFusionPass::Patterns()
311{311{
312+ if (!IsSupportedPlatform()) {
313+ return {};
314+ }
312 std::vector<PatternUniqPtr> patterns;315 std::vector<PatternUniqPtr> patterns;
313 patterns.reserve(kCommConfigNum * 2U);316 patterns.reserve(kCommConfigNum * 2U);
314 for (size_t config = 0U; config < kCommConfigNum; ++config) {317 for (size_t config = 0U; config < kCommConfigNum; ++config) {
@@ -326,9 +329,6 @@ std::vector<PatternUniqPtr> LayerNormFusionPass::Patterns()
326bool LayerNormFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)329bool LayerNormFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)
327{330{
328 OPS_LOG_D(kPassName.c_str(), "guard_begin.");331 OPS_LOG_D(kPassName.c_str(), "guard_begin.");
329- if (!IsSupportedPlatform()) {
330- return false;
331- }
332 Matched matched;332 Matched matched;
333 if (!CollectMatched(match_result, matched)) {333 if (!CollectMatched(match_result, matched)) {
334 return false;334 return false;
Mnorm/layer_norm/op_graph/fusion_pass/layer_norm_onnx_fusion_pass.cpp+3-3
@@ -635,6 +635,9 @@ bool CreateAffineConst(es::EsGraphBuilder& builder, const Matched& m, es::EsTens
635std::vector<PatternUniqPtr> LayerNormONNXFusionPass::Patterns()635std::vector<PatternUniqPtr> LayerNormONNXFusionPass::Patterns()
636{636{
637 OPS_LOG_D(kPassName.c_str(), "Enter Patterns for LayerNormONNXFusionPass.");637 OPS_LOG_D(kPassName.c_str(), "Enter Patterns for LayerNormONNXFusionPass.");
638+ if (!IsSupportedPlatform()) {
639+ return {};
640+ }
638 std::vector<PatternUniqPtr> patterns;641 std::vector<PatternUniqPtr> patterns;
639 patterns.reserve(kExpectedPatternNum);642 patterns.reserve(kExpectedPatternNum);
640 for (const auto& shape : kShapes) {643 for (const auto& shape : kShapes) {
@@ -651,9 +654,6 @@ std::vector<PatternUniqPtr> LayerNormONNXFusionPass::Patterns()
651bool LayerNormONNXFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)654bool LayerNormONNXFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)
652{655{
653 OPS_LOG_D(kPassName.c_str(), "guard_begin.");656 OPS_LOG_D(kPassName.c_str(), "guard_begin.");
654- if (!IsSupportedPlatform()) {
655- return false;
656- }
657 Matched m;657 Matched m;
658 if (!CollectMatched(match_result, m)) {658 if (!CollectMatched(match_result, m)) {
659 return false;659 return false;
Mnorm/layer_norm/op_graph/fusion_pass/layer_norm_onnx_mul_fusion_pass.cpp+3-3
@@ -532,6 +532,9 @@ bool CreateAffineConst(es::EsGraphBuilder& builder, const Matched& m, es::EsTens
532std::vector<PatternUniqPtr> LayerNormONNXMULFusionPass::Patterns()532std::vector<PatternUniqPtr> LayerNormONNXMULFusionPass::Patterns()
533{533{
534 OPS_LOG_D(kPassName.c_str(), "Enter Patterns for LayerNormONNXMULFusionPass.");534 OPS_LOG_D(kPassName.c_str(), "Enter Patterns for LayerNormONNXMULFusionPass.");
535+ if (!IsSupportedPlatform()) {
536+ return {};
537+ }
535 std::vector<PatternUniqPtr> patterns;538 std::vector<PatternUniqPtr> patterns;
536 patterns.reserve(kExpectedPatternNum);539 patterns.reserve(kExpectedPatternNum);
537 for (const bool with_affine : {true, false}) {540 for (const bool with_affine : {true, false}) {
@@ -548,9 +551,6 @@ std::vector<PatternUniqPtr> LayerNormONNXMULFusionPass::Patterns()
548bool LayerNormONNXMULFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)551bool LayerNormONNXMULFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)
549{552{
550 OPS_LOG_D(kPassName.c_str(), "guard_begin.");553 OPS_LOG_D(kPassName.c_str(), "guard_begin.");
551- if (!IsSupportedPlatform()) {
552- return false;
553- }
554 Matched m;554 Matched m;
555 if (!CollectMatched(match_result, m)) {555 if (!CollectMatched(match_result, m)) {
556 return false;556 return false;
Mnorm/layer_norm/op_graph/fusion_pass/layer_norm_remove_broadcast_fusion_pass.cpp+4-4
@@ -214,6 +214,10 @@ std::vector<es::EsTensorHolder> CreateReplacementInputs(es::EsGraphBuilder& grap
214 214 
215std::vector<PatternUniqPtr> LayerNormRemoveBroadcastFusionPass::Patterns()215std::vector<PatternUniqPtr> LayerNormRemoveBroadcastFusionPass::Patterns()
216{216{
217+ if (!IsSupportedPlatform()) {
218+ OPS_LOG_D(kPassName.c_str(), "Platform is not support.");
219+ return {};
220+ }
217 std::vector<PatternUniqPtr> patterns;221 std::vector<PatternUniqPtr> patterns;
218 patterns.emplace_back(MakePattern());222 patterns.emplace_back(MakePattern());
219 return patterns;223 return patterns;
@@ -222,10 +226,6 @@ std::vector<PatternUniqPtr> LayerNormRemoveBroadcastFusionPass::Patterns()
222bool LayerNormRemoveBroadcastFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)226bool LayerNormRemoveBroadcastFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)
223{227{
224 OPS_LOG_D(kPassName.c_str(), "Enter LayerNormRemoveBroadcastFusionPass MeetRequirements.");228 OPS_LOG_D(kPassName.c_str(), "Enter LayerNormRemoveBroadcastFusionPass MeetRequirements.");
225- if (!IsSupportedPlatform()) {
226- OPS_LOG_D(kPassName.c_str(), "Platform is not support.");
227- return false;
228- }
229 229 
230 GNode ln_node;230 GNode ln_node;
231 if (!GetCapturedNode(match_result, kLNCaptureIdx, ln_node)) {231 if (!GetCapturedNode(match_result, kLNCaptureIdx, ln_node)) {
Mnorm/layer_norm/op_graph/fusion_pass/layernorm_inference_fusion_pass.cpp+3-3
@@ -262,6 +262,9 @@ PatternUniqPtr MakePattern(size_t config)
262 262 
263std::vector<PatternUniqPtr> LayerNormInferenceFusionPass::Patterns()263std::vector<PatternUniqPtr> LayerNormInferenceFusionPass::Patterns()
264{264{
265+ if (!IsSupportedPlatform()) {
266+ return {};
267+ }
265 std::vector<PatternUniqPtr> patterns;268 std::vector<PatternUniqPtr> patterns;
266 patterns.reserve(kCommConfigNum);269 patterns.reserve(kCommConfigNum);
267 for (size_t config = 0U; config < kCommConfigNum; ++config) {270 for (size_t config = 0U; config < kCommConfigNum; ++config) {
@@ -277,9 +280,6 @@ std::vector<PatternUniqPtr> LayerNormInferenceFusionPass::Patterns()
277bool LayerNormInferenceFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)280bool LayerNormInferenceFusionPass::MeetRequirements(const std::unique_ptr<MatchResult>& match_result)
278{281{
279 OPS_LOG_D(kPassName.c_str(), "guard_begin.");282 OPS_LOG_D(kPassName.c_str(), "guard_begin.");
280- if (!IsSupportedPlatform()) {
281- return false;
282- }
283 Matched matched;283 Matched matched;
284 if (!CollectMatched(match_result, matched)) {284 if (!CollectMatched(match_result, matched)) {
285 return false;285 return false;
Mnorm/layer_norm_v4/docs/aclnnFastLayerNorm.md+2-2
@@ -159,7 +159,7 @@ aclnnStatus aclnnFastLayerNorm(
159 <td>meanOutOptional（aclTensor*）</td>159 <td>meanOutOptional（aclTensor*）</td>
160 <td>输出</td>160 <td>输出</td>
161 <td>可选输出，表示进行归一化后的均值。对应公式中的`meanOutOptional`。</td>161 <td>可选输出，表示进行归一化后的均值。对应公式中的`meanOutOptional`。</td>
162- <td><ul><li>支持空Tensor。</li><li>数据类型与`input`的数据类型保持一致。</li><li>shape为[A1,...,Ai,1,...,1]，Ai后共有j个1，与需要norm的轴长度保持相同。当`rstdOutOptional`存在时，shape与`rstdOutOptional`相同。</li></ul></td>162+ <td><ul><li>支持空Tensor。</li><li>shape为[A1,...,Ai,1,...,1]，Ai后共有j个1，与需要norm的轴长度保持相同。当`rstdOutOptional`存在时，shape与`rstdOutOptional`相同。</li></ul></td>
163 <td>FLOAT32、FLOAT16、BFLOAT16</td>163 <td>FLOAT32、FLOAT16、BFLOAT16</td>
164 <td>ND</td>164 <td>ND</td>
165 <td>1-8</td>165 <td>1-8</td>
@@ -169,7 +169,7 @@ aclnnStatus aclnnFastLayerNorm(
169 <td>rstdOutOptional（aclTensor*）</td>169 <td>rstdOutOptional（aclTensor*）</td>
170 <td>输出</td>170 <td>输出</td>
171 <td>可选输出，表示进行归一化后的标准差倒数。对应公式中的`rstdOutOptional`。</td>171 <td>可选输出，表示进行归一化后的标准差倒数。对应公式中的`rstdOutOptional`。</td>
172- <td><ul><li>支持空Tensor。</li><li>数据类型与`input`的数据类型保持一致。</li><li>shape为[A1,...,Ai,1,...,1]，Ai后共有j个1，与需要norm的轴长度保持相同。当`meanOutOptional`存在时，shape与`meanOutOptional`相同。</li></ul></td>172+ <td><ul><li>支持空Tensor。</li><li>shape为[A1,...,Ai,1,...,1]，Ai后共有j个1，与需要norm的轴长度保持相同。当`meanOutOptional`存在时，shape与`meanOutOptional`相同。</li></ul></td>
173 <td>FLOAT32、FLOAT16、BFLOAT16</td>173 <td>FLOAT32、FLOAT16、BFLOAT16</td>
174 <td>ND</td>174 <td>ND</td>
175 <td>1-8</td>175 <td>1-8</td>