Add post ops to ClGemmMatrixMultiplyReshapedOnlyRHSKernel and ClGemmMatrixMultiplyNativeKernel Part 3
Partially resolves: COMPMID-4435
Change-Id: Ifc5affa3a24a70942ca2d001380205df09b03ad7
Signed-off-by: SiCongLi <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6550
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/framework/Macros.h b/tests/framework/Macros.h
index a6ba137..ac03bb0 100644
--- a/tests/framework/Macros.h
+++ b/tests/framework/Macros.h
@@ -49,8 +49,8 @@
#define CONCAT(ARG0, ARG1) ARG0##ARG1
-#define VARIADIC_SIZE_IMPL(e0, e1, e2, e3, e4, e5, e6, e7, e8, e9, e10, e11, e12, e13, size, ...) size
-#define VARIADIC_SIZE(...) VARIADIC_SIZE_IMPL(__VA_ARGS__, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
+#define VARIADIC_SIZE_IMPL(e0, e1, e2, e3, e4, e5, e6, e7, e8, e9, e10, e11, e12, e13, e14, e15, size, ...) size
+#define VARIADIC_SIZE(...) VARIADIC_SIZE_IMPL(__VA_ARGS__, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
#define JOIN_PARAM1(OP, param) OP(0, param)
#define JOIN_PARAM2(OP, param, ...) \
@@ -92,6 +92,12 @@
#define JOIN_PARAM14(OP, param, ...) \
OP(13, param) \
, JOIN_PARAM13(OP, __VA_ARGS__)
+#define JOIN_PARAM15(OP, param, ...) \
+ OP(14, param) \
+ , JOIN_PARAM14(OP, __VA_ARGS__)
+#define JOIN_PARAM16(OP, param, ...) \
+ OP(15, param) \
+ , JOIN_PARAM15(OP, __VA_ARGS__)
#define JOIN_PARAM(OP, NUM, ...) \
CONCAT(JOIN_PARAM, NUM) \
(OP, __VA_ARGS__)
diff --git a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
index dc5fbc3..e3f151a 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
@@ -53,6 +53,11 @@
template <typename T>
using CLGEMMMatrixMultiplyNativeFixture = GEMMMatrixMultiplyNativeValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
+// Fixture for CLGEMMMatrixMultiplyNative with post ops
+template <typename T>
+using CLGEMMMatrixMultiplyNativeWithPostOpsFixture =
+ GEMMMatrixMultiplyNativeWithPostOpsValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
+
// Fixture for CLGEMMMatrixMultiplyNative3D
template <typename T>
using CLGEMMMatrixMultiplyNative3DFixture = GEMMMatrixMultiplyNative3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
@@ -141,6 +146,80 @@
broadcast_bias_values),
framework::dataset::make("Activation", ActivationLayerInfo()));
+/** Post Ops */
+using PostOpArgBroadcast = CLGEMMMatrixMultiplyNativeWithPostOpsFixture<float>::PostOpArgBroadcast;
+experimental::PostOpList<PostOpArgBroadcast> post_ops_1()
+{
+ experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+ post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F});
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
+ std::make_tuple(true, true, false), // If broadcast in dims 0, 1 and 2
+ 0,
+ ConvertPolicy::SATURATE);
+ post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
+ return post_ops;
+}
+experimental::PostOpList<PostOpArgBroadcast> post_ops_2()
+{
+ experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
+ std::make_tuple(false, true, true), // If broadcast in dims 0, 1 and 2
+ 1,
+ ConvertPolicy::SATURATE);
+ post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
+ return post_ops;
+}
+experimental::PostOpList<PostOpArgBroadcast> post_ops_3()
+{
+ experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+ // post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
+ std::make_tuple(false, false, false), // If broadcast in dims 0, 1 and 2
+ 1,
+ ConvertPolicy::SATURATE);
+ return post_ops;
+}
+
+/** Different Post Op Lists */
+const auto post_op_lists = framework::dataset::make("post_op_lists", {
+ post_ops_1(),
+ post_ops_2(),
+ post_ops_3(),
+} );
+
+bool is_post_op_list_valid(unsigned int m, unsigned int n, unsigned int k, unsigned int batch, DataType data_type, const experimental::PostOpList<ITensorInfo*>& post_ops)
+{
+ const auto lhs_info = GEMMLHSMatrixInfo(4,4,1,false,true);
+ const auto rhs_info = GEMMRHSMatrixInfo(4,4,1,true,true,false);
+
+ // Create TensorInfo for post op arguments
+ TensorInfo input0_info(TensorShape(k, m, batch), 1, data_type);
+ TensorInfo input1_info(TensorShape(n, k, batch), 1, data_type);
+ TensorInfo input2_info(TensorShape(n), 1, data_type);
+ TensorInfo output_info(TensorShape(n, m, batch), 1, data_type);
+
+ GEMMKernelInfo gemm_info(m, n, k, 0 /**< Depth of the output tensor in case is reinterpreted as 3D */,
+ false /**< reinterpret the input as 3D */,
+ true /**< Flag used to broadcast the bias addition */,
+ false /**< wider accumm */,
+ false /**< has pad y */,
+ ActivationLayerInfo::ActivationFunction::IDENTITY,
+ 1 /**< Multiplication factor for the width of the 1xW transposed block */,
+ 1 /**< Multiplication factor for the height of the 4x4 interleaved block */,
+ lhs_info,
+ rhs_info,
+ 0 /**< Offset to be added to each element of the matrix A */,
+ 0 /**< Offset to be added to each element of the matrix B */,
+ post_ops);
+ return bool(ClGemmMatrixMultiplyNativeKernel::validate(&input0_info.clone()->set_is_resizable(true),
+ &input1_info.clone()->set_is_resizable(true),
+ &input2_info.clone()->set_is_resizable(true),
+ &output_info.clone()->set_is_resizable(true),1.f,1.f,
+ lhs_info,
+ rhs_info,
+ gemm_info));
+}
+
/** 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, const ActivationLayerInfo &act_info)
{
@@ -191,6 +270,119 @@
TEST_SUITE(CL)
TEST_SUITE(GEMMMatrixMultiplyNative)
+TEST_SUITE(ValidateFusedPostOpsConfigs)
+TEST_SUITE(Invalid)
+TEST_CASE(UnsupportedPostOpSequence, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 17;
+ const unsigned int n = 1;
+ const unsigned int k = 13;
+ const unsigned int batch = 2;
+ TensorShape post_op_arg0_shape(n, m, batch);
+ TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type);
+ auto post_op_arg1_info = post_op_arg_info.clone();
+
+ // Unsupported sequence of post ops
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>(
+ &post_op_arg_info,
+ 1,
+ ConvertPolicy::SATURATE);
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>(
+ post_op_arg1_info.get(),
+ 0,
+ ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS);
+}
+TEST_CASE(OutputWidened, framework::DatasetMode::ALL)
+{
+ // Invalid broadcast: post op tensors "widen" the output tensor
+ const auto data_type = DataType::F32;
+ const unsigned int m = 1;
+ const unsigned int n = 18;
+ const unsigned int k = 13;
+ const unsigned int batch = 2;
+ TensorShape post_op_arg_shape(n, m + 1, batch); // output's Y dimension (m) is "widened", which is not allowed
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS);
+}
+TEST_CASE(BroadcastInXDimOnly, framework::DatasetMode::ALL)
+{
+ // Invalid broadcast: post op tensors broadcast in the first dimension (X) only
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ TensorShape post_op_arg_shape(1, m, batch);
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS);
+}
+TEST_SUITE_END() // Invalid
+TEST_SUITE(Valid)
+TEST_CASE(EmptyPostOpList, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
+TEST_CASE(BroadcastInYDimOnly, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ TensorShape post_op_arg_shape(n, 1, batch);
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
+TEST_CASE(BroadcastInBothXandYDims, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ TensorShape post_op_arg_shape(1, 1, batch);
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
+TEST_CASE(BroadcastInAllDims, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ TensorShape post_op_arg_shape(1, 1, 1);
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
+TEST_SUITE_END() // Valid
+TEST_SUITE_END() // ValidateFusedPostOps
TEST_SUITE(Float)
TEST_SUITE(FP32)
DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(
@@ -323,6 +515,32 @@
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
+
+TEST_SUITE(FusedPostOps)
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeWithPostOpsFixture<float>, framework::DatasetMode::ALL,
+ combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+ m_values,
+ n_values),
+ k_values),
+ b_values),
+ framework::dataset::make("M0", { 4 })),
+ n0_values_precommit),
+ k0_values_precommit),
+ framework::dataset::make("DataType", DataType::F32)),
+ framework::dataset::make("alpha", {1.0f} )),
+ framework::dataset::make("beta", {1.0f} )),
+ framework::dataset::make("broadcast_bias", { false, true } )),
+ framework::dataset::make("Activation", { ActivationLayerInfo() })),
+ post_op_lists)
+ )
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+TEST_SUITE_END() // FusedPostOps
+
TEST_SUITE_END() // FP32
TEST_SUITE_END() // Float
TEST_SUITE_END() // GEMMMatrixMulipltyNative
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
index 0f86a70..9e1a185 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
@@ -26,6 +26,7 @@
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/runtime/CL/CLTensor.h"
#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "src/core/experimental/PostOp.h"
#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h"
#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h"
#include "tests/CL/CLAccessor.h"
@@ -61,6 +62,11 @@
template <typename T>
using CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture = GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshapedOnlyRHS>;
+// Fixture for CLGEMMMatrixMultiplyReshapedOnlyRHS with post ops
+template <typename T>
+using CLGEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsFixture =
+ GEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshapedOnlyRHS>;
+
namespace
{
// *INDENT-OFF*
@@ -157,6 +163,81 @@
broadcast_bias_values),
framework::dataset::make("Activation", ActivationLayerInfo()));
+/** Post Ops */
+using PostOpArgBroadcast = CLGEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsFixture<float>::PostOpArgBroadcast;
+experimental::PostOpList<PostOpArgBroadcast> post_ops_1()
+{
+ experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+ post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F});
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
+ std::make_tuple(true, true, false), // If broadcast in dims 0, 1 and 2
+ 0,
+ ConvertPolicy::SATURATE);
+ post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
+ return post_ops;
+}
+experimental::PostOpList<PostOpArgBroadcast> post_ops_2()
+{
+ experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
+ std::make_tuple(false, true, true), // If broadcast in dims 0, 1 and 2
+ 1,
+ ConvertPolicy::SATURATE);
+ post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
+ return post_ops;
+}
+experimental::PostOpList<PostOpArgBroadcast> post_ops_3()
+{
+ experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+ post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
+ std::make_tuple(false, false, true), // If broadcast in dims 0, 1 and 2
+ 1,
+ ConvertPolicy::SATURATE);
+ return post_ops;
+}
+
+/** Different Post Op Lists */
+const auto post_op_lists = framework::dataset::make("post_op_lists", {
+ post_ops_1(),
+ post_ops_2(),
+ post_ops_3(),
+ } );
+
+ bool is_post_op_list_valid(unsigned int m, unsigned int n, unsigned int k, unsigned int batch, DataType data_type, const experimental::PostOpList<ITensorInfo*>& post_ops)
+{
+ const auto lhs_info = GEMMLHSMatrixInfo(4,4,1,false,true);
+ const auto rhs_info = GEMMRHSMatrixInfo(4,4,1,true,true,false);
+
+ // Create TensorInfo for post op arguments
+ TensorInfo input0_info(TensorShape(k, m, batch), 1, data_type);
+ TensorInfo input1_info(TensorShape(n, k, batch), 1, data_type);
+ TensorInfo input2_info(TensorShape(n), 1, data_type);
+ TensorInfo output_info(TensorShape(n, m, batch), 1, data_type);
+
+ const TensorInfo reshaped_input1_info = input1_info.clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(input1_info, rhs_info));
+
+ GEMMKernelInfo gemm_info(m, n, k, 0 /**< Depth of the output tensor in case is reinterpreted as 3D */,
+ false /**< reinterpret the input as 3D */,
+ true /**< Flag used to broadcast the bias addition */,
+ false /**< wider accumm */,
+ false /**< has pad y */,
+ ActivationLayerInfo::ActivationFunction::IDENTITY,
+ 1 /**< Multiplication factor for the width of the 1xW transposed block */,
+ 1 /**< Multiplication factor for the height of the 4x4 interleaved block */,
+ lhs_info,
+ rhs_info,
+ 0 /**< Offset to be added to each element of the matrix A */,
+ 0 /**< Offset to be added to each element of the matrix B */,
+ post_ops);
+ return bool(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(&input0_info.clone()->set_is_resizable(true),
+ &reshaped_input1_info.clone()->set_is_resizable(true),
+ &input2_info.clone()->set_is_resizable(true),
+ &output_info.clone()->set_is_resizable(true),1.f,1.f,
+ lhs_info,
+ rhs_info,
+ gemm_info));
+}
/** Configuration test */
bool 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,
@@ -211,6 +292,7 @@
CLGEMMMatrixMultiplyReshapedOnlyRHS gemm;
return bool(gemm.validate(&lhs, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info));
}
+
} // namespace
TEST_SUITE(CL)
@@ -262,6 +344,119 @@
ARM_COMPUTE_EXPECT(status == expected_value, framework::LogLevel::ERRORS);
}
+TEST_SUITE(ValidateFusedPostOpsConfigs)
+TEST_SUITE(Invalid)
+TEST_CASE(UnsupportedPostOpSequence, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 17;
+ const unsigned int n = 1;
+ const unsigned int k = 13;
+ const unsigned int batch = 2;
+ TensorShape post_op_arg0_shape(n, m, batch);
+ TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type);
+ auto post_op_arg1_info = post_op_arg_info.clone();
+
+ // Unsupported sequence of post ops
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>(
+ &post_op_arg_info,
+ 1,
+ ConvertPolicy::SATURATE);
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>(
+ post_op_arg1_info.get(),
+ 0,
+ ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS);
+}
+TEST_CASE(OutputWidened, framework::DatasetMode::ALL)
+{
+ // Invalid broadcast: post op tensors "widen" the output tensor
+ const auto data_type = DataType::F32;
+ const unsigned int m = 17;
+ const unsigned int n = 1;
+ const unsigned int k = 1;
+ const unsigned int batch = 1;
+ TensorShape post_op_arg_shape(n, m, batch + 4); // output's batch dimension is "widened", which is not allowed
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS);
+}
+TEST_CASE(BroadcastInXDimOnly, framework::DatasetMode::ALL)
+{
+ // Invalid broadcast: post op tensors broadcast in the first dimension (X) only
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ TensorShape post_op_arg_shape(1, m, batch);
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS);
+}
+TEST_SUITE_END() // Invalid
+TEST_SUITE(Valid)
+TEST_CASE(EmptyPostOpList, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
+TEST_CASE(BroadcastInYDimOnly, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ TensorShape post_op_arg_shape(n, 1, batch);
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
+TEST_CASE(BroadcastInBothXandYDims, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ TensorShape post_op_arg_shape(1, 1, batch);
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
+TEST_CASE(BroadcastInAllDims, framework::DatasetMode::ALL)
+{
+ const auto data_type = DataType::F32;
+ const unsigned int m = 22;
+ const unsigned int n = 16;
+ const unsigned int k = 15;
+ const unsigned int batch = 3;
+ TensorShape post_op_arg_shape(1, 1, 1);
+ TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+ experimental::PostOpList<ITensorInfo*> post_ops{};
+ post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
+
+ ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
+TEST_SUITE_END() // Valid
+TEST_SUITE_END() // ValidateFusedPostOps
TEST_SUITE(Float)
TEST_SUITE(FP32)
@@ -462,6 +657,44 @@
framework::ARM_COMPUTE_PRINT_INFO();
}
}
+
+TEST_SUITE(FusedPostOps)
+
+FIXTURE_DATA_TEST_CASE(RunPrecommit, CLGEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsFixture<float>, framework::DatasetMode::ALL,
+ combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+ m_values,
+ n_values),
+ k_values),
+ b_values),
+ m0_values_precommit),
+ n0_values_precommit),
+ k0_values_precommit),
+ framework::dataset::make("H0", {1})),
+ framework::dataset::make("interleave_rhs", { true })),
+ t_values_rhs),
+ framework::dataset::make("export_to_cl_image_rhs", false, true)),
+ framework::dataset::make("DataType", DataType::F32)),
+ a_values),
+ beta_values),
+ framework::dataset::make("broadcast_bias", { false } )),
+ act_values),
+ post_op_lists)
+ )
+{
+ // Validate output only if the target platform supports the OpenCL cl_khr_image2d_from_buffer extension
+ if(validate_result)
+ {
+ validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+ }
+ else
+ {
+ ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped");
+ framework::ARM_COMPUTE_PRINT_INFO();
+ }
+}
+
+TEST_SUITE_END() // FusedPostOps
+
TEST_SUITE_END() // FP32
TEST_SUITE(FP16)
@@ -590,6 +823,43 @@
framework::ARM_COMPUTE_PRINT_INFO();
}
}
+TEST_SUITE(FusedPostOps)
+
+FIXTURE_DATA_TEST_CASE(RunPrecommit, CLGEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsFixture<half>, framework::DatasetMode::ALL,
+ combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+ m_values,
+ n_values),
+ k_values),
+ b_values),
+ m0_values_precommit),
+ n0_values_precommit),
+ k0_values_precommit),
+ framework::dataset::make("H0", {1})),
+ framework::dataset::make("interleave_rhs", { true })),
+ t_values_rhs),
+ framework::dataset::make("export_to_cl_image_rhs", true)),
+ framework::dataset::make("DataType", DataType::F16)),
+ a_values),
+ beta_values),
+ framework::dataset::make("broadcast_bias", { false } )),
+ act_values),
+ post_op_lists)
+ )
+{
+ // Validate output only if the target platform supports the OpenCL cl_khr_image2d_from_buffer extension
+ if(validate_result)
+ {
+ validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16);
+ }
+ else
+ {
+ ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped");
+ framework::ARM_COMPUTE_PRINT_INFO();
+ }
+}
+
+TEST_SUITE_END() // FusedPostOps
+
TEST_SUITE_END() // FP16
TEST_SUITE_END() // Float
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index e119158..fa27301 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -1522,6 +1522,243 @@
SimpleTensor<T> _reference{};
};
+/** (EXPERIMENTAL_POST_OPS)*/
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType>
+class GEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsValidationFixture : public framework::Fixture
+{
+public:
+ using PostOpArgBroadcast = std::tuple<bool, bool, bool>; // Instruct fixture if we need broadcasting in dimension 0, 1, 2 of each PostOp argument
+ 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, bool export_to_cl_image, DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info,
+ const experimental::PostOpList<PostOpArgBroadcast> &post_ops)
+ {
+ GEMMLHSMatrixInfo lhs_info;
+ lhs_info.m0 = m0;
+ lhs_info.k0 = k0;
+
+ GEMMRHSMatrixInfo rhs_info;
+ rhs_info.n0 = n0;
+ rhs_info.k0 = k0;
+ rhs_info.h0 = h0;
+ rhs_info.interleave = interleave_rhs;
+ rhs_info.transpose = transpose_rhs;
+ rhs_info.export_to_cl_image = export_to_cl_image;
+
+ // 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);
+ auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops,
+ [ = ](auto broadcast)
+ {
+ return TensorShape
+ {
+ std::get<0>(broadcast) ? 1 : n,
+ std::get<1>(broadcast) ? 1 : m,
+ std::get<2>(broadcast) ? 1 : batch_size,
+ };
+ });
+
+ _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes);
+ if(validate_result)
+ {
+ _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes);
+ }
+ }
+
+protected:
+ template <typename U>
+ void fill(U &&tensor, int i)
+ {
+ static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported.");
+ using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type;
+
+ DistributionType distribution{ T(-1.0f), T(1.0f) };
+ library->fill(tensor, distribution, i);
+
+ // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0)
+ DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) };
+ library->fill_borders_with_garbage(tensor, distribution_inf, i);
+ }
+
+ TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+ DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops)
+ {
+ // Create tensors
+ 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;
+ // Create post op tensors and populate post op with them
+ std::vector<TensorType> post_op_tensors_holder{};
+ auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, ITensorInfo *>(post_ops,
+ [&post_op_tensors_holder, &data_type](auto shape)
+ {
+ auto t = create_tensor<TensorType>(shape, data_type, 1);
+ post_op_tensors_holder.push_back(std::move(t));
+ return post_op_tensors_holder.back().info();
+ });
+
+ const unsigned int M = lhs_shape[1];
+ const unsigned int N = rhs_shape[0];
+ const unsigned int K = lhs_shape[0];
+ GEMMKernelInfo kernel_info;
+ kernel_info.m = M;
+ kernel_info.n = N;
+ kernel_info.k = K;
+ 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;
+ kernel_info.post_ops = populated_post_ops;
+
+ // The output tensor will be auto-initialized within the function
+
+ // Create and configure function
+ ReshapeRHSOperatorType reshape_rhs;
+ GEMMOperatorType gemm;
+
+ validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info));
+ validate_result = validate_result || !rhs_info.export_to_cl_image;
+ if(!validate_result)
+ {
+ return nullptr;
+ }
+
+ reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info);
+ gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info);
+
+ ARM_COMPUTE_ASSERT(lhs.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(rhs.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(bias.info()->is_resizable());
+ for(const auto &tensor : post_op_tensors_holder)
+ {
+ ARM_COMPUTE_ASSERT(tensor.info()->is_resizable());
+ }
+
+ // We do not pad when using image as it needs to comply to strict pitch alignment restrictions
+ if(!rhs_info.export_to_cl_image)
+ {
+ add_padding_x({ &lhs, &rhs, &rhs_reshaped, &bias, &dst });
+ for(auto &tensor : post_op_tensors_holder)
+ {
+ add_padding_x({ &tensor });
+ }
+ }
+
+ // Allocate tensors
+ lhs.allocator()->allocate();
+ rhs.allocator()->allocate();
+ rhs_reshaped.allocator()->allocate();
+ bias.allocator()->allocate();
+ dst.allocator()->allocate();
+ for(auto &tensor : post_op_tensors_holder)
+ {
+ tensor.allocator()->allocate();
+ }
+
+ ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(!bias.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(!dst.info()->is_resizable());
+ for(const auto &tensor : post_op_tensors_holder)
+ {
+ ARM_COMPUTE_ASSERT(!tensor.info()->is_resizable());
+ }
+
+ // Fill tensors
+ fill(AccessorType(lhs), 0);
+ fill(AccessorType(rhs), 1);
+ fill(AccessorType(bias), 2);
+ for(size_t i = 0; i < post_op_tensors_holder.size(); ++i)
+ {
+ fill(AccessorType(post_op_tensors_holder.at(i)), 3 + i);
+ }
+
+ // Compute GEMM
+ ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } };
+ reshape_rhs.run(reshape_rhs_pack);
+ ITensorPack gemm_pack({ { ACL_SRC_0, &lhs },
+ { ACL_SRC_1, &rhs_reshaped },
+ { ACL_SRC_2, &bias },
+ { ACL_DST, &dst }
+ });
+ for(size_t i = 0; i < post_op_tensors_holder.size(); ++i)
+ {
+ gemm_pack.add_tensor(experimental::get_post_op_arg_type(i), &post_op_tensors_holder.at(i));
+ }
+ gemm.run(gemm_pack);
+
+ return dst;
+ }
+
+ SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias,
+ const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops)
+ {
+ TensorShape dst_shape = lhs_shape;
+ dst_shape[0] = rhs_shape[0];
+ dst_shape[1] = lhs_shape[1];
+
+ // Create reference
+ SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+ SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+ SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+ // Create post op tensors and populate post op with them
+ auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, SimpleTensor<T>>(post_ops, [&data_type](auto shape)
+ {
+ return SimpleTensor<T> { 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
+ int tensor_idx = 0;
+ fill(lhs, tensor_idx++);
+ fill(rhs, tensor_idx++);
+ fill(bias, tensor_idx++);
+ for(auto &op : populated_post_ops.get_list())
+ {
+ for(auto tensor : op->arguments())
+ {
+ fill(*tensor, tensor_idx++);
+ }
+ }
+
+ 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));
+ }
+ }
+
+ SimpleTensor<T> out;
+ out = reference::gemm<T>(lhs, rhs, bias, alpha, beta);
+ // Ignore activation info if post ops are used instead
+ if(populated_post_ops.size() > 0)
+ {
+ out = reference::post_ops<T>(out, populated_post_ops);
+ }
+ else
+ {
+ out = reference::activation_layer(out, act_info);
+ }
+ return out;
+ }
+
+ bool validate_result = true;
+ TensorType _target{};
+ SimpleTensor<T> _reference{};
+};
+
template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType>
class GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture
{
@@ -1830,6 +2067,213 @@
};
template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType>
+class GEMMMatrixMultiplyNativeWithPostOpsValidationFixture : public framework::Fixture
+{
+public:
+ using PostOpArgBroadcast = std::tuple<bool, bool, bool>; // Instruct fixture if we need broadcasting in dimension 0, 1, 2 of each PostOp argument
+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,
+ const ActivationLayerInfo &act_info, const experimental::PostOpList<PostOpArgBroadcast> &post_ops)
+ {
+ GEMMLHSMatrixInfo lhs_info;
+ lhs_info.m0 = m0;
+ lhs_info.k0 = k0;
+
+ GEMMRHSMatrixInfo rhs_info;
+ rhs_info.n0 = n0;
+ rhs_info.k0 = k0;
+
+ // 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);
+ const auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops,
+ [ = ](auto broadcast)
+ {
+ return TensorShape
+ {
+ std::get<0>(broadcast) ? 1 : n,
+ std::get<1>(broadcast) ? 1 : m,
+ std::get<2>(broadcast) ? 1 : batch_size,
+ };
+ });
+
+ _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes);
+ _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes);
+ }
+
+protected:
+ template <typename U>
+ void fill(U &&tensor, int i)
+ {
+ static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported.");
+ using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type;
+
+ DistributionType distribution{ T(-1.0f), T(1.0f) };
+ library->fill(tensor, distribution, i);
+
+ // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0)
+ DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) };
+ library->fill_borders_with_garbage(tensor, distribution_inf, i);
+ }
+
+ TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+ DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops)
+ {
+ // Create tensors
+ 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;
+ // Create post op tensors and populate post op with them
+ std::vector<TensorType> post_op_tensors_holder{};
+ auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, ITensorInfo *>(post_ops,
+ [&post_op_tensors_holder, &data_type](auto shape)
+ {
+ auto t = create_tensor<TensorType>(shape, data_type, 1);
+ post_op_tensors_holder.push_back(std::move(t));
+ return post_op_tensors_holder.back().info();
+ });
+
+ const unsigned int M = lhs_shape[1];
+ const unsigned int N = rhs_shape[0];
+ const unsigned int K = lhs_shape[0];
+ GEMMKernelInfo kernel_info;
+ kernel_info.m = M;
+ kernel_info.n = N;
+ kernel_info.k = K;
+ 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;
+ kernel_info.post_ops = populated_post_ops;
+
+ // Create and configure function
+ GEMMOperatorType gemm;
+ gemm.configure(lhs.info(), rhs.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info);
+
+ ARM_COMPUTE_ASSERT(lhs.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(rhs.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(bias.info()->is_resizable());
+ for(const auto &tensor : post_op_tensors_holder)
+ {
+ ARM_COMPUTE_ASSERT(tensor.info()->is_resizable());
+ }
+
+ add_padding_x({ &lhs, &rhs, &bias, &dst });
+ for(auto &tensor : post_op_tensors_holder)
+ {
+ add_padding_x({ &tensor });
+ }
+
+ // Allocate tensors
+ lhs.allocator()->allocate();
+ rhs.allocator()->allocate();
+ bias.allocator()->allocate();
+ dst.allocator()->allocate();
+ for(auto &tensor : post_op_tensors_holder)
+ {
+ tensor.allocator()->allocate();
+ }
+
+ ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(!bias.info()->is_resizable());
+ ARM_COMPUTE_ASSERT(!dst.info()->is_resizable());
+ for(const auto &tensor : post_op_tensors_holder)
+ {
+ ARM_COMPUTE_ASSERT(!tensor.info()->is_resizable());
+ }
+
+ // Fill tensors
+ fill(AccessorType(lhs), 0);
+ fill(AccessorType(rhs), 1);
+ fill(AccessorType(bias), 2);
+ for(size_t i = 0; i < post_op_tensors_holder.size(); ++i)
+ {
+ fill(AccessorType(post_op_tensors_holder.at(i)), 3 + i);
+ }
+
+ // Compute GEMM
+ ITensorPack gemm_pack({ { ACL_SRC_0, &lhs },
+ { ACL_SRC_1, &rhs },
+ { ACL_SRC_2, &bias },
+ { ACL_DST, &dst }
+ });
+ for(size_t i = 0; i < post_op_tensors_holder.size(); ++i)
+ {
+ gemm_pack.add_tensor(experimental::get_post_op_arg_type(i), &post_op_tensors_holder.at(i));
+ }
+ gemm.run(gemm_pack);
+
+ return dst;
+ }
+
+ SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias,
+ const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops)
+ {
+ TensorShape dst_shape = lhs_shape;
+ dst_shape[0] = rhs_shape[0];
+ dst_shape[1] = lhs_shape[1];
+
+ // Create reference
+ SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+ SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+ SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+ // Create post op tensors and populate post op with them
+ auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, SimpleTensor<T>>(post_ops, [&data_type](auto shape)
+ {
+ return SimpleTensor<T> { 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
+ int tensor_idx = 0;
+ fill(lhs, tensor_idx++);
+ fill(rhs, tensor_idx++);
+ fill(bias, tensor_idx++);
+ for(auto &op : populated_post_ops.get_list())
+ {
+ for(auto tensor : op->arguments())
+ {
+ fill(*tensor, tensor_idx++);
+ }
+ }
+
+ 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));
+ }
+ }
+
+ SimpleTensor<T> out;
+ out = reference::gemm<T>(lhs, rhs, bias, alpha, beta);
+ // Ignore activation info if post ops are used instead
+ if(populated_post_ops.size() > 0)
+ {
+ out = reference::post_ops<T>(out, populated_post_ops);
+ }
+ else
+ {
+ out = reference::activation_layer(out, act_info);
+ }
+ return out;
+ }
+
+ TensorType _target{};
+ SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType>
class GEMMMatrixMultiplyNative3DValidationFixture : public framework::Fixture
{
public: