COMPMID-2171: Fuse bias addition with CLGEMMMatrixMultiplyReshapedOnlyRHSKernel

Change-Id: I1d1e1f28fe7022309d72900893e8368820ca0f89
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1259
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 119 insertions, 43 deletions

diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
index 83051d2efe..23ae004912 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
@@ -72,6 +72,9 @@ constexpr float         tolerance_num_f16 = 0.02f;
 /** Alpha values to test - Precommit */
 const auto a_values = framework::dataset::make("alpha", {1.0f, -0.75f} );
 
+/** Beta values to test - Precommit */
+const auto beta_values = framework::dataset::make("beta", {-0.75f, 0.0f} );
+
 /** M values to test */
 const auto m_values = framework::dataset::make("M", 37);
 
@@ -120,8 +123,11 @@ const auto i_values_rhs = framework::dataset::make("interleave_rhs", { true, fal
 /** Transpose values to test with RHS matrix */
 const auto t_values_rhs = framework::dataset::make("transpose_rhs", { true, false });
 
+/**Broadcast bias from vector to matrix */
+const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", {false, true} );
+
 /** Configuration test */
-void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, unsigned int h0_value, bool i_value_rhs, bool t_value_rhs, DataType data_type)
+void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, unsigned int h0_value, bool i_value_rhs, bool t_value_rhs, bool broadcast_bias, DataType data_type)
 {
     const unsigned int M = m_value;
     const unsigned int N = n_value;
@@ -138,7 +144,7 @@ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned
     rhs_info.interleave = i_value_rhs;
     rhs_info.transpose  = t_value_rhs;
 
-    GEMMReshapeInfo gemm_info(M, N, K);
+    GEMMReshapeInfo gemm_info(M, N, K, false, false, 0, false, broadcast_bias);
 
     const TensorShape lhs_shape(K, M, b_value);
     const TensorShape rhs_shape(N, K, b_value);
@@ -154,13 +160,22 @@ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned
     CLTensor rhs_reshaped = create_tensor<CLTensor>(rhs_shape_reshaped, data_type);
     CLTensor dst          = create_tensor<CLTensor>(dst_shape, data_type);
 
+    TensorShape bias_shape = dst_shape;
+    if (broadcast_bias)
+    {
+        bias_shape[1] = 1;
+        bias_shape[2] = 1;
+    }
+    CLTensor bias         = create_tensor<CLTensor>(bias_shape, data_type);
+
     ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
 
     // Create and configure function
     CLGEMMMatrixMultiplyReshapedOnlyRHS gemm;
-    gemm.configure(&lhs, &rhs_reshaped, &dst, 1.0f, lhs_info, rhs_info, gemm_info);
+    gemm.configure(&lhs, &rhs_reshaped, &bias, &dst, 1.0f, 1.0f, lhs_info, rhs_info, gemm_info);
 }
 } // namespace
 
@@ -168,7 +183,7 @@ TEST_SUITE(CL)
 TEST_SUITE(GEMMMatrixMultiplyReshapedOnlyRHS)
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -179,13 +194,14 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi
                                                                    h0_values_precommit),
                                                                    i_values_rhs),
                                                                    t_values_rhs),
-m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, h0_value, i_value_rhs, t_value_rhs)
+                                                                   broadcast_bias_values),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, h0_value, i_value_rhs, t_value_rhs, broadcast_bias)
 {
-    validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, h0_value, i_value_rhs, t_value_rhs, DataType::F32);
+    validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, h0_value, i_value_rhs, t_value_rhs, broadcast_bias, DataType::F32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<float>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -197,14 +213,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<floa
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<float>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -216,14 +234,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<floa
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<float>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_w_values,
                                                                    m_h_values),
                                                                    n_values),
@@ -236,14 +256,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   b_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<float>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_w_values,
                                                                    m_h_values),
                                                                    n_values),
@@ -256,7 +277,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   b_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
@@ -265,7 +287,7 @@ TEST_SUITE_END() // FP32
 
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<half>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -277,14 +299,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<half
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   b_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<half>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -296,14 +320,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<half
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   b_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<half>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_w_values,
                                                                    m_h_values),
                                                                    n_values),
@@ -316,14 +342,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   b_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<half>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_w_values,
                                                                    m_h_values),
                                                                    n_values),
@@ -336,7 +363,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   b_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
@@ -347,4 +375,4 @@ TEST_SUITE_END() // GEMMMatrixMulipltyReshapedOnlyRHS
 TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index b7976104aa..34f9bd848c 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -390,7 +390,7 @@ class GEMMMatrixMultiplyReshapedOnlyRHSValidationFixture : public framework::Fix
 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)
+               bool interleave_rhs, bool transpose_rhs, DataType data_type, float alpha, float beta, bool broadcast_bias)
     {
         GEMMLHSMatrixInfo lhs_info;
         lhs_info.m0 = m0;
@@ -407,8 +407,18 @@ public:
         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);
+        TensorShape bias_shape;
+        if(broadcast_bias)
+        {
+            bias_shape = TensorShape(n, 1, 1);
+        }
+        else
+        {
+            bias_shape = TensorShape(n, m, batch_size);
+        }
+
+        _target    = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias);
+        _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias);
     }
 
 protected:
@@ -423,11 +433,13 @@ protected:
         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)
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                              DataType data_type, float alpha, float beta, bool broadcast_bias)
     {
         // Create tensors
-        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
-        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType lhs  = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs  = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
         TensorType rhs_reshaped;
         TensorType dst;
 
@@ -441,7 +453,7 @@ protected:
         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));
+        gemm.configure(&lhs, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, 0, false, broadcast_bias));
 
         ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -450,6 +462,7 @@ protected:
         lhs.allocator()->allocate();
         rhs.allocator()->allocate();
         rhs_reshaped.allocator()->allocate();
+        bias.allocator()->allocate();
         dst.allocator()->allocate();
 
         ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -460,6 +473,7 @@ protected:
         // Fill tensors
         fill(AccessorType(lhs), 0);
         fill(AccessorType(rhs), 1);
+        fill(AccessorType(bias), 2);
 
         // Compute GEMM
         reshape_rhs.run();
@@ -468,7 +482,7 @@ protected:
         return dst;
     }
 
-    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha)
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias)
     {
         TensorShape dst_shape = lhs_shape;
         dst_shape[0]          = rhs_shape[0];
@@ -477,13 +491,31 @@ protected:
         // 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 };
+        SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+
+        const int n          = rhs_shape[0];
+        const int m          = lhs_shape[1];
+        const int batch_size = lhs_shape[2];
 
         // Fill reference
         fill(lhs, 0);
         fill(rhs, 1);
 
-        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+        if(broadcast_bias)
+        {
+            SimpleTensor<T> tmp{ bias_shape, data_type, 1 };
+            fill(tmp, 2);
+            for(int i = 0; i < m * batch_size; i++)
+            {
+                memcpy(bias.data() + i * n, tmp.data(), n * sizeof(T));
+            }
+        }
+        else
+        {
+            fill(bias, 2);
+        }
+
+        return (reference::gemm<T>(lhs, rhs, bias, alpha, beta));
     }
 
     TensorType      _target{};
@@ -590,7 +622,7 @@ class GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::F
 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)
+               bool interleave_rhs, bool transpose_rhs, DataType data_type, float alpha, float beta)
     {
         GEMMLHSMatrixInfo lhs_info;
         lhs_info.m0 = m0;
@@ -609,9 +641,10 @@ public:
         // 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);
+        const TensorShape bias_shape(n, 1, 1);
 
-        _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);
+        _target    = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h);
+        _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h);
     }
 
 protected:
@@ -622,12 +655,14 @@ protected:
         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,
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                              DataType data_type, float alpha, float beta,
                               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 lhs  = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs  = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
         TensorType rhs_reshaped;
         TensorType dst;
 
@@ -641,7 +676,7 @@ protected:
         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));
+        gemm.configure(&lhs, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h, false, true));
 
         ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -650,6 +685,7 @@ protected:
         lhs.allocator()->allocate();
         rhs.allocator()->allocate();
         rhs_reshaped.allocator()->allocate();
+        bias.allocator()->allocate();
         dst.allocator()->allocate();
 
         ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -660,6 +696,7 @@ protected:
         // Fill tensors
         fill(AccessorType(lhs), 0);
         fill(AccessorType(rhs), 1);
+        fill(AccessorType(bias), 2);
 
         // Compute GEMM
         reshape_rhs.run();
@@ -668,7 +705,7 @@ protected:
         return dst;
     }
 
-    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, unsigned int m_h)
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h)
     {
         TensorShape dst_shape = lhs_shape;
         dst_shape.set(0, rhs_shape[0]);
@@ -679,13 +716,24 @@ protected:
         // 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 };
+        SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+
+        const int n          = rhs_shape[0];
+        const int m          = lhs_shape[1];
+        const int batch_size = lhs_shape[2];
 
         // Fill reference
         fill(lhs, 0);
         fill(rhs, 1);
 
-        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+        SimpleTensor<T> tmp{ bias_shape, data_type, 1 };
+        fill(tmp, 2);
+        for(int i = 0; i < m * batch_size; i++)
+        {
+            memcpy(bias.data() + i * n, tmp.data(), n * sizeof(T));
+        }
+
+        return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
     }
 
     TensorType      _target{};