IVGCVSW-6494 Add CpuAcc Batch MatMul Workload Fp32

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I2def6995f81d33e68f1ea45d8d19a1e6294049b1

3 files changed, 39 insertions, 156 deletions

diff --git a/src/backends/backendsCommon/WorkloadUtils.cpp b/src/backends/backendsCommon/WorkloadUtils.cpp
index b045530abc..3aea667bfe 100644
--- a/src/backends/backendsCommon/WorkloadUtils.cpp
+++ b/src/backends/backendsCommon/WorkloadUtils.cpp
@@ -341,4 +341,24 @@ std::map<std::string, unsigned int> CalculateGatherNdKeyIndices(TensorInfo input
     return keyIndices;
 }
 
+armnn::PermutationVector GeneratePermutationVectorOnLastTwoDimensions(unsigned int rank)
+{
+    armnn::PermutationVector permutationVector{};
+    switch (rank)
+    {
+        case 2:
+            permutationVector = {1U, 0U};
+            break;
+        case 3:
+            permutationVector = {0U, 2U, 1U};
+            break;
+        case 4:
+            permutationVector = {0U, 1U, 3U, 2U};
+            break;
+        default:
+            throw Exception("Invalid number of dimensions.");
+    }
+    return permutationVector;
+}
+
 } // namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadUtils.hpp b/src/backends/backendsCommon/WorkloadUtils.hpp
index 0e5487336f..3d8d927345 100644
--- a/src/backends/backendsCommon/WorkloadUtils.hpp
+++ b/src/backends/backendsCommon/WorkloadUtils.hpp
@@ -258,4 +258,10 @@ std::tuple<ConstTensor, unsigned int> Convert1HWOtoMIHW(const ConstTensorHandle*
 /// \return - A map with names and values for  N, ND, K, W, C
 std::map<std::string, unsigned int> CalculateGatherNdKeyIndices(TensorInfo inputInfo0, TensorInfo inputInfo1);
 
+/// Generates a permutation vector of size rank that permutes the 2 most right dimensions
+///
+/// \param rank - Tensor rank, i.e. number of dimensions in the tensors
+/// \return - A permutation vector that permutes the 2 last dimensions
+armnn::PermutationVector GeneratePermutationVectorOnLastTwoDimensions(unsigned int rank);
+
 }  //namespace armnn
diff --git a/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.cpp
index 6fcc35ab52..74bd97f103 100644
--- a/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.cpp
+++ b/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.cpp
@@ -71,20 +71,9 @@ LayerTestResult<T, 2> BatchMatMul2DSimpleTest(
 {
     auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({2,2}, ArmnnType, qScale, qOffset);
@@ -160,20 +149,9 @@ LayerTestResult<T, 3> BatchMatMul3DSimpleTest(
 {
     auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({1,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({1,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({1,2,2}, ArmnnType, qScale, qOffset);
@@ -249,20 +227,9 @@ LayerTestResult<T, 4> BatchMatMulNCHWSimpleTest(
 {
     auto descriptor = armnn::BatchMatMulDescriptor(); // Default arbitrary layout is treated the same as NCHW
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({1,1,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({1,1,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({1,1,2,2}, ArmnnType, qScale, qOffset);
@@ -343,20 +310,9 @@ LayerTestResult<T, 4> BatchMatMulNHWCSimpleTest(
                                                    armnn::DataLayout::NHWC,
                                                    armnn::DataLayout::NHWC);
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({1,2,2,1}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({1,2,2,1}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({1,2,2,1}, ArmnnType, qScale, qOffset);
@@ -432,20 +388,9 @@ LayerTestResult<T, 3> BatchMatMul3DBatchTest(
 {
     auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({2,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset);
@@ -530,20 +475,9 @@ LayerTestResult<T, 3> BatchMatMul3DBroadcastTest(
 {
     auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({1,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset);
@@ -625,20 +559,9 @@ LayerTestResult<T, 3> BatchMatMul3D2DBroadcastTest(
 {
     auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset);
@@ -725,20 +648,9 @@ LayerTestResult<T, 5> BatchMatMulNDHWCNHWCTest(
                                                    armnn::DataLayout::NDHWC,
                                                    armnn::DataLayout::NHWC);
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({1,1,2,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({1,2,2,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({1,1,2,2,2}, ArmnnType, qScale, qOffset);
@@ -823,20 +735,9 @@ LayerTestResult<T, 2> BatchMatMul2DTinyTest(
 {
     auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({1,1}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({1,1}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({1,1}, ArmnnType, qScale, qOffset);
@@ -909,20 +810,9 @@ LayerTestResult<T, 3> BatchMatMul3DNonSquareTest(
 {
     auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({2,5,3}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({2,3,4}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({2,5,4}, ArmnnType, qScale, qOffset);
@@ -1024,20 +914,9 @@ LayerTestResult<T, 2> BatchMatMul2DTranspSimpleTest(
                                                    false,
                                                    false);
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({2,3}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({2,3}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({3,3}, ArmnnType, qScale, qOffset);
@@ -1117,20 +996,9 @@ LayerTestResult<T, 2> BatchMatMul2DAdjointSimpleTest(
                                                    true,
                                                    false);
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({3,3}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({3,3}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({3,3}, ArmnnType, qScale, qOffset);
@@ -1227,20 +1095,9 @@ LayerTestResult<T, 4> BatchMatMulNHWCParamsTest(
                                                    armnn::DataLayout::NHWC,
                                                    armnn::DataLayout::NHWC);
 
-    float qScale = 0.0f;
+    float qScale = 1.0f;
     int32_t qOffset = 0;
 
-    switch(ArmnnType)
-    {
-        case armnn::DataType::QAsymmS8:
-        case armnn::DataType::QAsymmU8:
-        case armnn::DataType::QSymmS16:
-            qScale = 1.0f;
-            break;
-        default:
-            break;
-    }
-
     armnn::TensorInfo inputXInfo({1,4,4,2}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo inputYInfo({2,2,4,1}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputInfo({2,4,2,2}, ArmnnType, qScale, qOffset);