commit 1fe6c8170ae2fe90b53fb71b7570aec9dfe75c45
author Teresa Charlin <teresa.charlinreyes@arm.com> Tue Nov 01 15:59:50 2022 +0000
committer TeresaARM <teresa.charlinreyes@arm.com> Mon May 08 13:16:01 2023 +0000
tree bbb846edda64445c1e033b182e5a079c8d5728d8
parent c52190a7e80cf238ba1d8630e5cc36ec7c7849e2

IVGCVSW-7307 Add CpuAcc Batch MatMul Workload

* Call dedicated MatMul kernel in ACL
* Add int8 tests
* Add int8 to documentation
* Force tensors to be dynamic (nonConst) as per request of ACL

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I992ae9aae1174214607bf29305f21cdeaf3fdc1b

src/backends/neon/workloads/NeonBatchMatMulWorkload.hpp

diff --git a/src/backends/neon/workloads/NeonBatchMatMulWorkload.hpp b/src/backends/neon/workloads/NeonBatchMatMulWorkload.hpp
index cb004d2..27144f2 100644
--- a/src/backends/neon/workloads/NeonBatchMatMulWorkload.hpp
+++ b/src/backends/neon/workloads/NeonBatchMatMulWorkload.hpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -7,35 +7,27 @@
 
 #include "NeonBaseWorkload.hpp"
 
-#include <arm_compute/runtime/IFunction.h>
-#include <arm_compute/runtime/Tensor.h>
-
-#include <memory>
+#include <arm_compute/runtime/NEON/functions/NEMatMul.h>
 
 namespace armnn
 {
-    arm_compute::Status NeonBatchMatMulValidate(const TensorInfo& inputX,
-                                                const TensorInfo& inputY,
-                                                const TensorInfo& output,
-                                                const BatchMatMulDescriptor& descriptor);
+    arm_compute::Status NeonBatchMatMulValidate(const TensorInfo& inputInfoX,
+                                                const TensorInfo& inputInfoY,
+                                                const TensorInfo& outputInfo,
+                                                const BatchMatMulDescriptor& descriptor,
+                                                const bool isFastMathEnabled,
+                                                const ActivationDescriptor* activationDescriptor);
+
 
     class NeonBatchMatMulWorkload : public NeonBaseWorkload<BatchMatMulQueueDescriptor>
     {
     public:
         NeonBatchMatMulWorkload(const BatchMatMulQueueDescriptor& descriptor,
-                                const WorkloadInfo& info);
+                                const WorkloadInfo& info,
+                                const bool isFastMathEnabled);
         virtual void Execute() const override;
 
     private:
-        // ACL layers required to fully form a Batch Mat Mul layer.
-        std::unique_ptr<arm_compute::IFunction> m_GEMMLayer;
-        std::unique_ptr<arm_compute::IFunction> m_PermuteLayerX;
-        std::unique_ptr<arm_compute::IFunction> m_PermuteLayerY;
-
-        // Additional ACL arm_compute::Tensors.
-        // Required to perform permutations.
-        arm_compute::Tensor m_PermutedTensorX;
-        arm_compute::Tensor m_PermutedTensorY;
-
+        mutable arm_compute::NEMatMul m_MatMulLayer;
     };
 } //namespace armnn