COMPMID-1979: Fuse Activation Function in CLGEMM - part 2
Fuse activation function in:
CLGEMMMatrixMultiplyNativeKernel
CLGEMMMatrixMultiplyReshapedKernel
CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
Change-Id: I033ace2bdc58903594c9f31175e4b23c4b559f6f
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1565
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
diff --git a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
index 031b807..33b01d8 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
@@ -90,6 +90,13 @@
/** Batch size values to test */
const auto b_values = framework::dataset::make("batch_size", 1, 3);
+/** Activation values to test */
+const auto act_values = framework::dataset::make("Activation",
+{
+ ActivationLayerInfo(),
+ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
+});
+
/** M0 values to test - Precommit */
const auto m0_values_precommit = framework::dataset::make("M0", {4, 6});
@@ -115,7 +122,7 @@
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, bool broadcast_bias, 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 ActivationLayerInfo &act_info)
{
const unsigned int M = m_value;
const unsigned int N = n_value;
@@ -130,10 +137,11 @@
rhs_info.k0 = k0_value;
GEMMKernelInfo kernel_info;
- kernel_info.m = M;
- kernel_info.n = N;
- kernel_info.k = K;
- kernel_info.broadcast_bias = broadcast_bias;
+ kernel_info.m = M;
+ kernel_info.n = N;
+ kernel_info.k = K;
+ kernel_info.broadcast_bias = broadcast_bias;
+ kernel_info.activation_info = act_info;
const TensorShape lhs_shape(K, M, b_value);
const TensorShape rhs_shape(N, K, b_value);
@@ -165,7 +173,7 @@
TEST_SUITE(GEMMMatrixMultiplyNative)
TEST_SUITE(Float)
TEST_SUITE(FP32)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -174,13 +182,14 @@
n0_values_precommit),
k0_values_precommit),
broadcast_bias_values),
-m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias)
+ act_values),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias, act_value)
{
- validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias, DataType::F32);
+ validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias, DataType::F32, act_value);
}
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(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -191,14 +200,15 @@
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_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(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -209,14 +219,15 @@
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_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(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -227,14 +238,15 @@
k0_values_precommit),
framework::dataset::make("DataType", DataType::F32)),
a_values),
- beta_values))
+ beta_values),
+ act_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(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -245,7 +257,8 @@
k0_values_nightly),
framework::dataset::make("DataType", DataType::F32)),
a_values),
- beta_values))
+ beta_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
@@ -254,7 +267,7 @@
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(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -265,14 +278,15 @@
framework::dataset::make("DataType", DataType::F16)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_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(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -283,14 +297,15 @@
framework::dataset::make("DataType", DataType::F16)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_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(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -301,14 +316,15 @@
k0_values_precommit),
framework::dataset::make("DataType", DataType::F16)),
a_values),
- beta_values))
+ beta_values),
+ act_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(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -319,7 +335,8 @@
k0_values_nightly),
framework::dataset::make("DataType", DataType::F16)),
a_values),
- beta_values))
+ beta_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
index bd70ddb..99af296 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
@@ -98,6 +98,13 @@
/** Batch size values to test */
const auto b_values = framework::dataset::make("batch_size", 1, 3);
+/** Activation values to test */
+const auto act_values = framework::dataset::make("Activation",
+{
+ ActivationLayerInfo(),
+ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
+});
+
/** M0 values to test - Precommit */
const auto m0_values_precommit = framework::dataset::make("M0", {4, 6});
@@ -138,7 +145,7 @@
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 v0_value, unsigned int h0_value, bool i_value_lhs, bool i_value_rhs, bool broadcast_bias, 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 v0_value, unsigned int h0_value, bool i_value_lhs, bool i_value_rhs, bool broadcast_bias, DataType data_type, const ActivationLayerInfo &act_info)
{
const unsigned int M = m_value;
const unsigned int N = n_value;
@@ -165,6 +172,7 @@
kernel_info.depth_output_gemm3d = 0;
kernel_info.reinterpret_input_as_3d = false;
kernel_info.broadcast_bias = broadcast_bias;
+ kernel_info.activation_info = act_info;
const TensorShape lhs_shape(K, M, b_value);
const TensorShape lhs_shape_reshaped = compute_lhs_reshaped_shape(TensorInfo(lhs_shape, 1, data_type),
@@ -204,7 +212,7 @@
TEST_SUITE(GEMMMatrixMultiplyReshaped)
TEST_SUITE(Float)
TEST_SUITE(FP32)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(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(combine(combine(
m_values,
n_values),
k_values),
@@ -217,13 +225,14 @@
i_values_lhs),
i_values_rhs),
broadcast_bias_values),
-m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, broadcast_bias)
+ act_values),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, broadcast_bias, act_value)
{
- validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, broadcast_bias, DataType::F32);
+ validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, broadcast_bias, DataType::F32, act_value);
}
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<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(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -238,14 +247,15 @@
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<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(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -260,14 +270,15 @@
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<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(combine(combine(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -282,14 +293,15 @@
i_values_rhs),
framework::dataset::make("DataType", DataType::F32)),
a_values),
- beta_values))
+ beta_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<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(combine(combine(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -304,7 +316,8 @@
i_values_rhs),
framework::dataset::make("DataType", DataType::F32)),
a_values),
- beta_values))
+ beta_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
@@ -313,7 +326,7 @@
TEST_SUITE(FP16)
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<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(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -328,14 +341,15 @@
framework::dataset::make("DataType", DataType::F16)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
}
FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<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(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -350,14 +364,15 @@
framework::dataset::make("DataType", DataType::F16)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
}
FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<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(combine(combine(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -372,14 +387,15 @@
i_values_rhs),
framework::dataset::make("DataType", DataType::F16)),
a_values),
- beta_values))
+ beta_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
}
FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<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(combine(combine(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -394,7 +410,8 @@
i_values_rhs),
framework::dataset::make("DataType", DataType::F16)),
a_values),
- beta_values))
+ beta_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
index 6c34fc8..dd993af 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
@@ -94,6 +94,13 @@
/** Batch size values to test */
const auto b_values = framework::dataset::make("batch_size", 1, 3);
+/** Activation values to test */
+const auto act_values = framework::dataset::make("Activation",
+{
+ ActivationLayerInfo(),
+ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
+});
+
/** M0 values to test - Precommit */
const auto m0_values_precommit = framework::dataset::make("M0", {4, 6});
@@ -128,7 +135,7 @@
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, bool broadcast_bias, 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 ActivationLayerInfo &act_info)
{
const unsigned int M = m_value;
const unsigned int N = n_value;
@@ -152,6 +159,7 @@
kernel_info.depth_output_gemm3d = 0;
kernel_info.reinterpret_input_as_3d = false;
kernel_info.broadcast_bias = broadcast_bias;
+ kernel_info.activation_info = act_info;
const TensorShape lhs_shape(K, M, b_value);
const TensorShape rhs_shape(N, K, b_value);
@@ -187,7 +195,7 @@
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(combine(
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -199,13 +207,14 @@
i_values_rhs),
t_values_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)
+ act_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, act_value)
{
- 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);
+ 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, act_value);
}
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(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -219,14 +228,15 @@
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_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(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -240,14 +250,15 @@
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_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(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -261,14 +272,15 @@
t_values_rhs),
framework::dataset::make("DataType", DataType::F32)),
a_values),
- beta_values))
+ beta_values),
+ act_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(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -282,7 +294,8 @@
t_values_rhs),
framework::dataset::make("DataType", DataType::F32)),
a_values),
- beta_values))
+ beta_values),
+ act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
@@ -291,7 +304,7 @@
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(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -305,14 +318,15 @@
framework::dataset::make("DataType", DataType::F16)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_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(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -326,14 +340,15 @@
framework::dataset::make("DataType", DataType::F16)),
a_values),
beta_values),
- broadcast_bias_values))
+ broadcast_bias_values),
+ act_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(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -347,14 +362,15 @@
t_values_rhs),
framework::dataset::make("DataType", DataType::F16)),
a_values),
- beta_values))
+ beta_values),
+ act_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(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -368,7 +384,8 @@
t_values_rhs),
framework::dataset::make("DataType", DataType::F16)),
a_values),
- beta_values))
+ beta_values),
+ act_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 a225c71..ac8ab2a 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -33,6 +33,7 @@
#include "tests/framework/Asserts.h"
#include "tests/framework/Fixture.h"
#include "tests/validation/Helpers.h"
+#include "tests/validation/reference/ActivationLayer.h"
#include "tests/validation/reference/GEMM.h"
#include <random>
@@ -158,7 +159,7 @@
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 v0, unsigned int h0, bool interleave_lhs,
- bool interleave_rhs, DataType data_type, float alpha, float beta, bool broadcast_bias)
+ bool interleave_rhs, DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info)
{
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0;
@@ -181,8 +182,8 @@
broadcast_bias ? 1 : m,
broadcast_bias ? 1 : 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);
+ _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info);
+ _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, act_info);
}
protected:
@@ -198,7 +199,7 @@
}
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)
+ DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info)
{
// Create tensors
TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
@@ -218,6 +219,7 @@
kernel_info.depth_output_gemm3d = 0;
kernel_info.reinterpret_input_as_3d = false;
kernel_info.broadcast_bias = broadcast_bias;
+ kernel_info.activation_info = act_info;
// The output tensor will be auto-initialized within the function
@@ -261,7 +263,8 @@
return dst;
}
- 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)
+ 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,
+ const ActivationLayerInfo &act_info)
{
TensorShape dst_shape = lhs_shape;
dst_shape[0] = rhs_shape[0];
@@ -290,7 +293,7 @@
}
}
- return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
+ return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
}
TensorType _target{};
@@ -304,7 +307,7 @@
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 v0, unsigned int h0,
bool interleave_lhs,
- bool interleave_rhs, DataType data_type, float alpha, float beta)
+ bool interleave_rhs, DataType data_type, float alpha, float beta, const ActivationLayerInfo &act_info)
{
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0;
@@ -328,8 +331,8 @@
const TensorShape rhs_shape(n, k, batch_size);
const TensorShape bias_shape(n, 1, 1);
- _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);
+ _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, act_info);
+ _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, act_info);
}
protected:
@@ -341,7 +344,7 @@
}
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)
+ DataType data_type, float alpha, float beta, unsigned int m_h, const ActivationLayerInfo &act_info)
{
// Create tensors
TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
@@ -361,6 +364,7 @@
kernel_info.depth_output_gemm3d = m_h;
kernel_info.reinterpret_input_as_3d = false;
kernel_info.broadcast_bias = true;
+ kernel_info.activation_info = act_info;
// The output tensor will be auto-initialized within the function
@@ -404,7 +408,8 @@
return dst;
}
- 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)
+ 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,
+ const ActivationLayerInfo &act_info)
{
TensorShape dst_shape = lhs_shape;
dst_shape.set(0, rhs_shape[0]);
@@ -432,7 +437,7 @@
memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
}
- return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
+ return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
}
TensorType _target{};
@@ -445,7 +450,7 @@
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, float beta, bool broadcast_bias)
+ bool interleave_rhs, bool transpose_rhs, DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info)
{
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0;
@@ -465,8 +470,8 @@
broadcast_bias ? 1 : m,
broadcast_bias ? 1 : 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);
+ _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info);
+ _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, act_info);
}
protected:
@@ -482,7 +487,7 @@
}
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)
+ DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info)
{
// Create tensors
TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
@@ -501,6 +506,7 @@
kernel_info.depth_output_gemm3d = 0;
kernel_info.reinterpret_input_as_3d = false;
kernel_info.broadcast_bias = broadcast_bias;
+ kernel_info.activation_info = act_info;
// The output tensor will be auto-initialized within the function
@@ -539,7 +545,8 @@
return dst;
}
- 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)
+ 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,
+ const ActivationLayerInfo &act_info)
{
TensorShape dst_shape = lhs_shape;
dst_shape[0] = rhs_shape[0];
@@ -568,7 +575,7 @@
}
}
- return (reference::gemm<T>(lhs, rhs, bias, alpha, beta));
+ return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
}
TensorType _target{};
@@ -581,7 +588,7 @@
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, float beta)
+ bool interleave_rhs, bool transpose_rhs, DataType data_type, float alpha, float beta, const ActivationLayerInfo &act_info)
{
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0;
@@ -602,8 +609,8 @@
const TensorShape rhs_shape(n, k, batch_size);
const TensorShape bias_shape(n, 1, 1);
- _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);
+ _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, act_info);
+ _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, act_info);
}
protected:
@@ -616,7 +623,7 @@
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)
+ unsigned int m_h, const ActivationLayerInfo &act_info)
{
// Create tensors
TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
@@ -635,6 +642,7 @@
kernel_info.depth_output_gemm3d = m_h;
kernel_info.reinterpret_input_as_3d = false;
kernel_info.broadcast_bias = true;
+ kernel_info.activation_info = act_info;
// The output tensor will be auto-initialized within the function
@@ -673,7 +681,8 @@
return dst;
}
- 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)
+ 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,
+ const ActivationLayerInfo &act_info)
{
TensorShape dst_shape = lhs_shape;
dst_shape.set(0, rhs_shape[0]);
@@ -701,7 +710,7 @@
memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
}
- return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
+ return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
}
TensorType _target{};
@@ -713,7 +722,8 @@
{
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, float beta, bool broadcast_bias)
+ 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,
+ const ActivationLayerInfo &act_info)
{
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0;
@@ -730,8 +740,8 @@
broadcast_bias ? 1 : m,
broadcast_bias ? 1 : 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);
+ _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info);
+ _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, act_info);
}
protected:
@@ -747,7 +757,7 @@
}
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)
+ DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info)
{
// Create tensors
TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
@@ -765,6 +775,7 @@
kernel_info.depth_output_gemm3d = 0;
kernel_info.reinterpret_input_as_3d = false;
kernel_info.broadcast_bias = broadcast_bias;
+ kernel_info.activation_info = act_info;
// Create and configure function
GEMMFunctionType gemm;
@@ -796,7 +807,8 @@
return dst;
}
- 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)
+ 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,
+ const ActivationLayerInfo &act_info)
{
TensorShape dst_shape = lhs_shape;
dst_shape[0] = rhs_shape[0];
@@ -825,7 +837,7 @@
}
}
- return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
+ return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
}
TensorType _target{};
@@ -837,7 +849,8 @@
{
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, float beta)
+ 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,
+ const ActivationLayerInfo &act_info)
{
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0;
@@ -855,8 +868,8 @@
const TensorShape rhs_shape(n, k, batch_size);
const TensorShape bias_shape(n, 1, 1);
- _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);
+ _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, act_info);
+ _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, act_info);
}
protected:
@@ -868,7 +881,7 @@
}
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)
+ DataType data_type, float alpha, float beta, unsigned int m_h, const ActivationLayerInfo &act_info)
{
// Create tensors
TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
@@ -886,6 +899,7 @@
kernel_info.depth_output_gemm3d = m_h;
kernel_info.reinterpret_input_as_3d = false;
kernel_info.broadcast_bias = true;
+ kernel_info.activation_info = act_info;
// The output tensor will be auto-initialized within the function
@@ -919,7 +933,8 @@
return dst;
}
- 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)
+ 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,
+ const ActivationLayerInfo &act_info)
{
TensorShape dst_shape = lhs_shape;
dst_shape.set(0, rhs_shape[0]);
@@ -947,7 +962,7 @@
memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
}
- return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
+ return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
}
TensorType _target{};