ACL EP GEMM improvements (#2780)

When it is posible we use a fully connected layer instead of the gemm implementation. This will let the library use the best implementation based on the input data.
microsoft · Jan 7, 2020 · fdc0106 · fdc0106
1 parent f22bffe
commit fdc0106
Showing 1 changed file with 20 additions and 38 deletions.
diff --git a/onnxruntime/core/providers/acl/math/gemm.h b/onnxruntime/core/providers/acl/math/gemm.h
@@ -21,6 +21,7 @@
 //NEON
 #include "arm_compute/runtime/NEON/functions/NEGEMM.h"
 #include "arm_compute/runtime/NEON/functions/NETranspose.h"
+#include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
 
 #undef GEMM_ACL
 #define CACHE_TRANSPOSED_DATA
@@ -29,7 +30,7 @@ namespace onnxruntime {
 namespace acl {
 
 typedef struct {
-  std::shared_ptr<arm_compute::NEGEMM> layer;
+  std::shared_ptr<arm_compute::IFunction> layer;
   std::shared_ptr<arm_compute::Tensor> a, b, c, d;
   std::shared_ptr<arm_compute::MemoryManagerOnDemand> mm_layer;
 } ACLNEGEMM;
@@ -51,7 +52,6 @@ class Gemm : public onnxruntime::Gemm<T> {
     ORT_ENFORCE(info.GetAttr<float>("beta", &beta_).IsOK());
   }
 
-#ifdef GEMM_ACL
   Status Compute(OpKernelContext* context) const override {
     const auto X = context->Input<Tensor>(0);
     const auto W = context->Input<Tensor>(1);
@@ -66,6 +66,8 @@ class Gemm : public onnxruntime::Gemm<T> {
     int64_t N = helper.N();
     auto Y = context->Output(0, TensorShape({M, N}));
 
+    bool FC = ((alpha_ == 1 && beta_ == 1) || (alpha_ == 1 && beta_ == 0));
+
     int64_t K = helper.K();
     LOGS_DEFAULT(VERBOSE) << "Gemm ACL:" << std::endl;
     if (X) LOGS_DEFAULT(VERBOSE) << "X " << X->Shape().ToString().c_str() << std::endl;
@@ -93,7 +95,7 @@ class Gemm : public onnxruntime::Gemm<T> {
       tGEMM.d->allocator()->init(arm_compute::TensorInfo(arm_compute::TensorShape(N, M), tGEMM.a->info()->format()));
 
       // transpose
-      if (trans_B_ == CblasTrans) {
+      if (!FC && trans_B_ == CblasTrans) {
         auto trans_layer = std::make_shared<arm_compute::NETranspose>();
         tGEMM.b->allocator()->init(arm_compute::TensorInfo(ACLTensorShape(W->Shape()), arm_compute::Format::F32));
 
@@ -113,13 +115,23 @@ class Gemm : public onnxruntime::Gemm<T> {
       }
 
       tGEMM.mm_layer = ACLCreateMemoryManager();
-      tGEMM.layer = std::make_shared<arm_compute::NEGEMM>(tGEMM.mm_layer);
 
-      // configure GEMM
-      tGEMM.layer->configure(tGEMM.a.get(), tGEMM.b.get(), tGEMM.c.get(), tGEMM.d.get(), alpha_, beta_, arm_compute::GEMMInfo());
+      if(FC) {
+        auto layer = std::make_shared<arm_compute::NEFullyConnectedLayer>(tGEMM.mm_layer);
+        layer->configure(tGEMM.a.get(), tGEMM.b.get(), (B != nullptr && beta_ != 0) ? tGEMM.c.get() : nullptr, tGEMM.d.get());
+        tGEMM.layer = std::move(layer);
+      } else {
+#ifdef GEMM_ACL
+        auto layer = std::make_shared<arm_compute::NEGEMM>(tGEMM.mm_layer);
+        layer->configure(tGEMM.a.get(), tGEMM.b.get(), (B != nullptr && beta_ != 0) ? tGEMM.c.get() : nullptr, tGEMM.d.get(), alpha_, beta_, arm_compute::GEMMInfo());
+        tGEMM.layer = std::move(layer);
+#else
+        return onnxruntime::Gemm<T>::Compute(context);
+#endif
+      }
 
       // non-transpose
-      if (trans_B_ != CblasTrans) {
+      if (FC || trans_B_ != CblasTrans) {
         const T* b_data = W->template Data<T>();
         ACLImportMemory(tGEMM.b->allocator(), (void*)b_data, W->Shape().Size() * 4);
       }
@@ -132,7 +144,7 @@ class Gemm : public onnxruntime::Gemm<T> {
       pGEMM = &it->second;
 
       // transpose
-      if (trans_B_ == CblasTrans) {
+      if (!FC && trans_B_ == CblasTrans) {
 #ifndef CACHE_TRANSPOSED_DATA
         auto trans_layer = std::make_shared<arm_compute::NETranspose>();
         pGEMM->b->allocator()->init(arm_compute::TensorInfo(ACLTensorShape(W->Shape()), arm_compute::Format::F32));
@@ -193,36 +205,6 @@ class Gemm : public onnxruntime::Gemm<T> {
     gemmLayers.erase(this);
   }
 
-#else
-
-  Status Compute(OpKernelContext* context) const override {
-    const auto X = context->Input<Tensor>(0);
-    const auto W = context->Input<Tensor>(1);
-    const auto B = context->Input<Tensor>(2);
-
-    GemmHelper helper(X->Shape(), trans_A_ != CblasNoTrans, W->Shape(), trans_B_ != CblasNoTrans, B->Shape());
-
-    if (!helper.State().IsOK())
-      return helper.State();
-
-    int64_t M = helper.M();
-    int64_t N = helper.N();
-    auto Y = context->Output(0, TensorShape({M, N}));
-
-    int64_t K = helper.K();
-    LOGS_DEFAULT(VERBOSE) << "Gemm CPU:" << std::endl;
-    if (X) LOGS_DEFAULT(VERBOSE) << "X " << X->Shape().ToString().c_str() << std::endl;
-    if (W) LOGS_DEFAULT(VERBOSE) << "W " << W->Shape().ToString().c_str() << std::endl;
-    if (B) LOGS_DEFAULT(VERBOSE) << "B " << B->Shape().ToString().c_str() << std::endl;
-    LOGS_DEFAULT(VERBOSE) << "Y " << Y->Shape().ToString().c_str() << std::endl;
-    LOGS_DEFAULT(VERBOSE) << "M " << (int)M << ", N " << (int)N << ", K " << (int)K << std::endl;
-    LOGS_DEFAULT(VERBOSE) << "Alfa " << alpha_ << ", Beta " << beta_ << std::endl;
-    LOGS_DEFAULT(VERBOSE) << std::endl;
-
-    return onnxruntime::Gemm<T>::Compute(context);
-  }
-#endif
-
  private:
   static thread_local std::map<OpKernel*, ACLNEGEMM> gemmLayers;