Add Fused Activation to OpenCL MatMul
- Added fused activation to MatMul function interface
- Added fused activation to CL backend
- Includes tests for supported Activation Functions in MatMul
Resolves: [COMPMID-6192]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: Ie103212b600b60699eaf6a6394d609e6e1f5aba6
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/522465
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9714
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/CL/MatMul.cpp b/tests/validation/CL/MatMul.cpp
index 6364b16..5a262a8 100644
--- a/tests/validation/CL/MatMul.cpp
+++ b/tests/validation/CL/MatMul.cpp
@@ -26,6 +26,7 @@
#include "arm_compute/runtime/CL/functions/CLMatMul.h"
#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/ActivationFunctionsDataset.h"
#include "tests/framework/DatasetModes.h"
#include "tests/framework/Macros.h"
#include "tests/framework/TestCase.h"
@@ -44,11 +45,13 @@
{
namespace
{
-RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for fp32 data type */
-constexpr float abs_tolerance_f32(0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp32 data type in case using relative tolerance fails because of small values */
-constexpr float abs_tolerance_f16(0.001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16 data type in case using relative tolerance fails because of small values */
-RelativeTolerance<half_float::half> tolerance_f16(half(0.01)); /**< Tolerance value for comparing reference's output against implementation's output for fp16 data type */
-constexpr AbsoluteTolerance<uint8_t> tolerance_quant(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
+RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for fp32 data type */
+constexpr float abs_tolerance_f32(
+ 0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp32 data type in case using relative tolerance fails because of small values */
+constexpr float abs_tolerance_f16(
+ 0.001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16 data type in case using relative tolerance fails because of small values */
+RelativeTolerance<half_float::half> tolerance_f16(half(0.01)); /**< Tolerance value for comparing reference's output against implementation's output for fp16 data type */
+constexpr AbsoluteTolerance<uint8_t> tolerance_quant(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
} // namespace
template <typename T>
@@ -57,25 +60,71 @@
template <typename T>
using CLQuantizedMatMulFixture = QuantizedMatMulValidationFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>;
+template <typename T>
+using CLMatMulActivationFixture = MatMulValidationWithActivationFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>;
+
+template <typename T>
+using CLMatMulActivationAlphaBetaFixture = MatMulValidationWithActivationAlphaBetaFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>;
+
+template <typename T>
+using CLQuantizedMatMulActivationFixture = QuantizedMatMulValidationWithActivationFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>;
+
+/* The main act functions matmul is expected to support */
+const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo",
+{
+ ActivationLayerInfo(),
+ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f),
+ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.75f, 0.25f),
+ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH)
+});
+
+const auto ActivationFunctionsQuantizedDataset = concat(concat(concat(
+ framework::dataset::make("ActivationInfo", ActivationLayerInfo()),
+ framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))),
+ framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f))),
+ framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.75f, 0.25f)));
+
+/* Larger activation functions dataset, used during some nightly tests. */
+const auto AllActivationsDataset = combine(datasets::ActivationFunctions(), framework::dataset::make("AlphaBeta", { 0.5f, 1.f }));
+
+const auto AllQuantizedActivationsDataset = combine(concat(datasets::ActivationFunctionsQuantized(),
+ framework::dataset::make("ActivationFunction", { ActivationLayerInfo::ActivationFunction::HARD_SWISH,
+ ActivationLayerInfo::ActivationFunction::LEAKY_RELU
+ })),
+ framework::dataset::make("AlphaBeta", { 0.5f, 1.f }));
+
TEST_SUITE(CL)
TEST_SUITE(MatMul)
TEST_SUITE(Float)
TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallMatMulDataset(),
- framework::dataset::make("TransposeA", { false, true })),
- framework::dataset::make("TransposeB", { false, true })),
- framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulActivationFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallMatMulDataset(),
+ framework::dataset::make("TransposeA", { false, true })),
+ framework::dataset::make("TransposeB", { false, true })),
+ framework::dataset::make("DataType", DataType::F32)),
+ ActivationFunctionsDataset))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, CLMatMulFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeMatMulDataset(),
- framework::dataset::make("TransposeA", { false, true })),
- framework::dataset::make("TransposeB", { false, true })),
- framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLMatMulActivationFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+ framework::dataset::make("TransposeA", { false, true })),
+ framework::dataset::make("TransposeB", { false, true })),
+ framework::dataset::make("DataType", DataType::F32)),
+ ActivationFunctionsDataset))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunAllActivations, CLMatMulActivationAlphaBetaFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::SmallerMatMulDataset(),
+ framework::dataset::make("TransposeA", { false })),
+ framework::dataset::make("TransposeB", { true })),
+ framework::dataset::make("DataType", DataType::F32)),
+ AllActivationsDataset))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
@@ -85,19 +134,21 @@
TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallMatMulDataset(),
- framework::dataset::make("TransposeA", { false, true })),
- framework::dataset::make("TransposeB", { false, true })),
- framework::dataset::make("DataType", DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulActivationFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallMatMulDataset(),
+ framework::dataset::make("TransposeA", { false, true })),
+ framework::dataset::make("TransposeB", { false, true })),
+ framework::dataset::make("DataType", DataType::F16)),
+ ActivationFunctionsDataset))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, CLMatMulFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeMatMulDataset(),
- framework::dataset::make("TransposeA", { false, true })),
- framework::dataset::make("TransposeB", { false, true })),
- framework::dataset::make("DataType", DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLMatMulActivationFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+ framework::dataset::make("TransposeA", { false, true })),
+ framework::dataset::make("TransposeB", { false, true })),
+ framework::dataset::make("DataType", DataType::F16)),
+ ActivationFunctionsDataset))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
@@ -110,32 +161,30 @@
TEST_SUITE(QASYMM8)
FIXTURE_DATA_TEST_CASE(RunSmall, CLQuantizedMatMulFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(
- datasets::SmallMatMulDataset(),
- framework::dataset::make("TransposeA", { false, true })),
- framework::dataset::make("TransposeB", { false, true })),
- framework::dataset::make("DataType", DataType::QASYMM8)),
- framework::dataset::make("ActivationInfo", { ActivationLayerInfo() })),
- framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
- framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 50, 1) })),
- framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 30, -1) })),
- framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) }))
-)
+ datasets::SmallMatMulDataset(),
+ framework::dataset::make("TransposeA", { false, true })),
+ framework::dataset::make("TransposeB", { false, true })),
+ framework::dataset::make("DataType", DataType::QASYMM8)),
+ ActivationFunctionsQuantizedDataset),
+ framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
+ framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 50, 1) })),
+ framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 30, -1) })),
+ framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
}
FIXTURE_DATA_TEST_CASE(RunLarge, CLQuantizedMatMulFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine(
- datasets::LargeMatMulDataset(),
- framework::dataset::make("TransposeA", { false, true })),
- framework::dataset::make("TransposeB", { false, true })),
- framework::dataset::make("DataType", DataType::QASYMM8)),
- framework::dataset::make("ActivationInfo", { ActivationLayerInfo() })),
- framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
- framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 100, 1) })),
- framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 200, -1) })),
- framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) }))
-)
+ datasets::LargeMatMulDataset(),
+ framework::dataset::make("TransposeA", { false, true })),
+ framework::dataset::make("TransposeB", { false, true })),
+ framework::dataset::make("DataType", DataType::QASYMM8)),
+ ActivationFunctionsQuantizedDataset),
+ framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
+ framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 100, 1) })),
+ framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 200, -1) })),
+ framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
@@ -146,32 +195,45 @@
TEST_SUITE(QASYMM8_SIGNED)
FIXTURE_DATA_TEST_CASE(RunSmall, CLQuantizedMatMulFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(
- datasets::SmallMatMulDataset(),
- framework::dataset::make("TransposeA", { false, true })),
- framework::dataset::make("TransposeB", { false, true })),
- framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
- framework::dataset::make("ActivationInfo", { ActivationLayerInfo() })),
- framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
- framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 50, 1) })),
- framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 30, -1) })),
- framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) }))
-)
+ datasets::SmallMatMulDataset(),
+ framework::dataset::make("TransposeA", { false, true })),
+ framework::dataset::make("TransposeB", { false, true })),
+ framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
+ ActivationFunctionsQuantizedDataset),
+ framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
+ framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 50, 1) })),
+ framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 30, -1) })),
+ framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
}
FIXTURE_DATA_TEST_CASE(RunLarge, CLQuantizedMatMulFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine(
- datasets::LargeMatMulDataset(),
- framework::dataset::make("TransposeA", { false, true })),
- framework::dataset::make("TransposeB", { false, true })),
- framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
- framework::dataset::make("ActivationInfo", { ActivationLayerInfo() })),
- framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
- framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 100, 1) })),
- framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 200, -1) })),
- framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) }))
-)
+ datasets::LargeMatMulDataset(),
+ framework::dataset::make("TransposeA", { false, true })),
+ framework::dataset::make("TransposeB", { false, true })),
+ framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
+ ActivationFunctionsQuantizedDataset),
+ framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
+ framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 100, 1) })),
+ framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 200, -1) })),
+ framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_quant);
+}
+
+FIXTURE_DATA_TEST_CASE(RunAllActivations, CLQuantizedMatMulActivationFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine(
+ datasets::LargeMatMulDataset(),
+ framework::dataset::make("TransposeA", { false })),
+ framework::dataset::make("TransposeB", { true })),
+ framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
+ AllQuantizedActivationsDataset),
+ framework::dataset::make("NumberOfExtraRuns", { 0, 1 })),
+ framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 100, 1) })),
+ framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 200, -1) })),
+ framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);