COMPMID-2172: Fuse bias addition with CLGEMMMatrixMultiplyNativeKernel

Change-Id: I714b92ec001fc71172719b67fb66d490538b6948
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1399
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 109 insertions, 50 deletions

diff --git a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
index c7c390353a..b0d1fd2ad1 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
@@ -68,6 +68,9 @@ 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);
 
@@ -107,8 +110,11 @@ const auto n0_values_nightly = framework::dataset::make("N0", { 2, 3, 4, 8 });
 /** K0 values to test - Nightly */
 const auto k0_values_nightly = framework::dataset::make("K0", { 2, 3, 4, 8 });
 
+/** 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, 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, bool broadcast_bias, DataType data_type)
 {
     const unsigned int M = m_value;
     const unsigned int N = n_value;
@@ -122,27 +128,30 @@ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned
     rhs_info.n0         = n0_value;
     rhs_info.k0         = k0_value;
 
-    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);
-
+    const TensorShape bias_shape(N,
+                                 broadcast_bias? 1 : M,
+                                 broadcast_bias? 1 : b_value);
     const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape, 1, data_type),
                                                    TensorInfo(rhs_shape, 1, data_type),
                                                    gemm_info);
-
     // Create tensors
-    CLTensor lhs          = create_tensor<CLTensor>(lhs_shape, data_type);
+    CLTensor lhs  = create_tensor<CLTensor>(lhs_shape, data_type);
     CLTensor rhs  = create_tensor<CLTensor>(rhs_shape, data_type);
-    CLTensor dst          = create_tensor<CLTensor>(dst_shape, data_type);
+    CLTensor bias = create_tensor<CLTensor>(bias_shape, data_type);
+    CLTensor dst  = create_tensor<CLTensor>(dst_shape, data_type);
 
     ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(rhs.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
     CLGEMMMatrixMultiplyNative gemm;
-    gemm.configure(&lhs, &rhs, &dst, 1.0f, lhs_info, rhs_info, gemm_info);
+    gemm.configure(&lhs, &rhs, &bias, &dst, 1.0f, 1.0f, lhs_info, rhs_info, gemm_info);
 }
 } // namespace
 
@@ -150,7 +159,7 @@ TEST_SUITE(CL)
 TEST_SUITE(GEMMMatrixMultiplyNative)
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -158,13 +167,14 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi
                                                                    m0_values_precommit),
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
-m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value)
+                                                                   broadcast_bias_values),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias)
 {
-    validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, DataType::F32);
+    validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias, DataType::F32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -173,14 +183,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, frame
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
                                                                    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, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -189,14 +201,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<float>, frame
                                                                    n0_values_nightly),
                                                                    k0_values_nightly),
                                                                    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, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::ALL,
-                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),
@@ -206,14 +220,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<float>, f
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::NIGHTLY,
-                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),
@@ -223,7 +238,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, f
                                                                    n0_values_nightly),
                                                                    k0_values_nightly),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
@@ -232,7 +248,7 @@ TEST_SUITE_END() // FP32
 
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -241,14 +257,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<half>, framew
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -257,14 +275,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<half>, framew
                                                                    n0_values_nightly),
                                                                    k0_values_nightly),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::ALL,
-                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),
@@ -274,14 +294,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<half>, fr
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::NIGHTLY,
-                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),
@@ -291,7 +312,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<half>, fr
                                                                    n0_values_nightly),
                                                                    k0_values_nightly),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index fcb41bb0ba..b721d841f7 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -684,7 +684,7 @@ class GEMMMatrixMultiplyNativeValidationFixture : 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, DataType data_type, float alpha)
+    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha, float beta, bool broadcast_bias)
     {
         GEMMLHSMatrixInfo lhs_info;
         lhs_info.m0 = m0;
@@ -697,9 +697,12 @@ 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,
+                                     broadcast_bias ? 1 : m,
+                                     broadcast_bias ? 1 : 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);
+        _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:
@@ -714,11 +717,13 @@ 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 dst;
 
         const unsigned int M = lhs_shape[1];
@@ -727,23 +732,27 @@ protected:
 
         // Create and configure function
         GEMMFunctionType gemm;
-        gemm.configure(&lhs, &rhs, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
+        gemm.configure(&lhs, &rhs, &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);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Allocate tensors
         lhs.allocator()->allocate();
         rhs.allocator()->allocate();
+        bias.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(!bias.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);
+        fill(AccessorType(bias), 2);
 
         // Compute GEMM
         gemm.run();
@@ -751,7 +760,7 @@ 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];
@@ -760,13 +769,27 @@ 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);
+        fill(bias, 2);
 
-        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+        if(broadcast_bias)
+        {
+            // In case of broadcast, we need simply copy the first into the following "M" ones
+            for(int i = 1; i < m * batch_size; i++)
+            {
+                memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
+            }
+        }
+
+        return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
     }
 
     TensorType      _target{};
@@ -778,7 +801,7 @@ class GEMMMatrixMultiplyNative3DValidationFixture : 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, DataType data_type, float alpha)
+    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, DataType data_type, float alpha, float beta)
     {
         GEMMLHSMatrixInfo lhs_info;
         lhs_info.m0 = m0;
@@ -794,9 +817,10 @@ 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:
@@ -807,13 +831,13 @@ 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,
-                              unsigned int m_h)
+    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 rhs_reshaped;
+        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 dst;
 
         const unsigned int M = lhs_shape[1];
@@ -824,25 +848,27 @@ protected:
 
         // Create and configure function
         GEMMFunctionType gemm;
-        gemm.configure(&lhs, &rhs, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h));
+        gemm.configure(&lhs, &rhs, &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);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Allocate tensors
         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);
         ARM_COMPUTE_EXPECT(!rhs.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);
 
         // Fill tensors
         fill(AccessorType(lhs), 0);
         fill(AccessorType(rhs), 1);
+        fill(AccessorType(bias), 2);
 
         // Compute GEMM
         gemm.run();
@@ -850,7 +876,7 @@ 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]);
@@ -861,13 +887,24 @@ 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);
+        fill(bias, 2);
 
-        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+        // In case of broadcast, we need simply copy the first into the following "M" ones
+        for(int i = 1; i < m * batch_size; i++)
+        {
+            memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
+        }
+
+        return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
     }
 
     TensorType      _target{};