COMPMID-2000: Implement CLGEMMMatrixMultiplyReshapedOnlyRHS - Transposed

Change-Id: I364c7ec5a43ad391a73429489802b0e679ee0c6e
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/732
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 214 insertions, 0 deletions

diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index a6a3b67785..77d2ca61fb 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -383,6 +383,220 @@ protected:
     TensorType      _target{};
     SimpleTensor<T> _reference{};
 };
+
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyReshapedOnlyRHSValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0,
+               bool interleave_rhs, bool transpose_rhs, DataType data_type, float alpha)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0 = m0;
+        lhs_info.k0 = k0;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = n0;
+        rhs_info.k0         = k0;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = interleave_rhs;
+        rhs_info.transpose  = transpose_rhs;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type, alpha);
+        _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+
+        // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0)
+        std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity());
+        library->fill_borders_with_garbage(tensor, distribution_inf, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type, float alpha)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType rhs_reshaped;
+        TensorType dst;
+
+        const unsigned int M = lhs_shape[1];
+        const unsigned int N = rhs_shape[0];
+        const unsigned int K = lhs_shape[0];
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(&lhs, &rhs_reshaped, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        rhs_reshaped.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(lhs), 0);
+        fill(AccessorType(rhs), 1);
+
+        // Compute GEMM
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape[0]          = rhs_shape[0];
+        dst_shape[1]          = lhs_shape[1];
+
+        // Create reference
+        SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> c{ dst_shape, data_type, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0,
+               bool interleave_rhs, bool transpose_rhs, DataType data_type, float alpha)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0 = m0;
+        lhs_info.k0 = k0;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = n0;
+        rhs_info.k0         = k0;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = interleave_rhs;
+        rhs_info.transpose  = transpose_rhs;
+
+        // In case of GEMM3D, m is the product between m_w and m_h
+        const unsigned int m = m_w * m_h;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type, alpha, m_h);
+        _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, m_h);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type, float alpha,
+                              unsigned int m_h)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType rhs_reshaped;
+        TensorType dst;
+
+        const unsigned int M = lhs_shape[1];
+        const unsigned int N = rhs_shape[0];
+        const unsigned int K = lhs_shape[0];
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(&lhs, &rhs_reshaped, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        rhs_reshaped.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(lhs), 0);
+        fill(AccessorType(rhs), 1);
+
+        // Compute GEMM
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, unsigned int m_h)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape.set(0, rhs_shape[0]);
+        dst_shape.set(1, lhs_shape[1] / m_h);
+        dst_shape.set(2, m_h);
+        dst_shape.set(3, lhs_shape[2]);
+
+        // Create reference
+        SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> c{ dst_shape, data_type, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute