Implement dynamic quantization for GEMMLowp tests
This patch calculates the output quantization info based on the inputs'
quantization information.
The previous approach was using the same quantization information for
input, weights and output.
Remove QSYMM8_PER_CHANNEL path from the fixture as there are no
related tests
Remove repeated shapes from the dataset now that we get rid of the
quantization info from the dataset.
Combine signed and unsigned SmallGEMMLowpFusedBatchedMatMulDataset
into one as they become identical
Resolves COMPMID-6481, COMPMID-6634
Change-Id: I9f5a20f4bb45c3e5adab388564135ae8a5c0a9ea
Signed-off-by: SiCong Li <sicong.li@arm.com>
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10680
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/CL/GEMMLowp.cpp b/tests/validation/CL/GEMMLowp.cpp
index 0b057b9..1ae9e96 100644
--- a/tests/validation/CL/GEMMLowp.cpp
+++ b/tests/validation/CL/GEMMLowp.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2022 Arm Limited.
+ * Copyright (c) 2017-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -44,6 +44,9 @@
{
namespace validation
{
+
+using framework::dataset::make;
+
namespace
{
constexpr AbsoluteTolerance<float> tolerance_quant(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
@@ -72,9 +75,9 @@
TEST_SUITE(BatchedMatMul)
TEST_SUITE(QASYMM8)
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedUnsigned, framework::DatasetMode::ALL,
- combine(combine(datasets::SmallGEMMLowpFusedBatchedMatMulDatasetUnsigned(),
- framework::dataset::make("DataType", { DataType::QASYMM8 })),
- framework::dataset::make("bool", { false })))
+ combine(datasets::SmallGEMMLowpFusedBatchedMatMulDataset(),
+ make("DataType", { DataType::QASYMM8 }),
+ make("reshape_b_only_on_first_run", { false })))
{
validate(CLAccessor(_target), _reference, tolerance_quant);
}
@@ -84,9 +87,9 @@
GEMMLowpMatrixMultiplyCoreFusedOffsetOutputGenericValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, false, int8_t, int8_t, true>;
TEST_SUITE(QASYMM8_SIGNED)
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedSigned, framework::DatasetMode::ALL,
- combine(combine(datasets::SmallGEMMLowpFusedBatchedMatMulDatasetSigned(),
- framework::dataset::make("DataType", { DataType::QASYMM8_SIGNED })),
- framework::dataset::make("bool", { false })))
+ combine(datasets::SmallGEMMLowpFusedBatchedMatMulDataset(),
+ make("DataType", { DataType::QASYMM8_SIGNED }),
+ make("reshape_b_only_on_first_run", { false })))
{
validate(CLAccessor(_target), _reference, tolerance_quant);
}
@@ -96,9 +99,10 @@
TEST_SUITE(FusedOffsetOutput)
TEST_SUITE(QASYMM8)
using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputGenericValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore>;
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture, framework::DatasetMode::ALL, combine(combine(datasets::SmallGEMMLowpFusedOffsetOutputUint8Dataset(),
- framework::dataset::make("DataType", { DataType::QASYMM8 })),
- framework::dataset::make("reshape_b_only_on_first_run", { true, false })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture, framework::DatasetMode::ALL,
+ combine(datasets::SmallGEMMLowpFusedOffsetOutputUint8Dataset(),
+ make("DataType", { DataType::QASYMM8 }),
+ make("reshape_b_only_on_first_run", { true, false })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
@@ -108,9 +112,9 @@
using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputOutput3DUint8Fixture =
GEMMLowpMatrixMultiplyCoreFusedOffsetOutputGenericValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, true>;
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputOutput3DUint8Fixture, framework::DatasetMode::ALL,
- combine(combine(datasets::SmallGEMMLowpFusedOffsetOutputOutput3DUint8Dataset(),
- framework::dataset::make("DataType", { DataType::QASYMM8 })),
- framework::dataset::make("reshape_b_only_on_first_run", { true, false })))
+ combine(datasets::SmallGEMMLowpFusedOffsetOutputOutput3DUint8Dataset(),
+ make("DataType", { DataType::QASYMM8 }),
+ make("reshape_b_only_on_first_run", { true, false })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
@@ -121,18 +125,19 @@
using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInputOutput3DUint8Fixture =
GEMMLowpMatrixMultiplyCoreFusedOffsetOutputGenericValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, true, true>;
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInputOutput3DUint8Fixture, framework::DatasetMode::ALL,
- combine(combine(datasets::SmallGEMMLowpFusedOffsetOutputInputOutput3DUint8Dataset(),
- framework::dataset::make("DataType", { DataType::QASYMM8 })),
- framework::dataset::make("reshape_b_only_on_first_run", { true, false })))
+ combine(datasets::SmallGEMMLowpFusedOffsetOutputInputOutput3DUint8Dataset(),
+ make("DataType", { DataType::QASYMM8 }),
+ make("reshape_b_only_on_first_run", { true, false })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
}
TEST_SUITE_END() // InputOutput3D
-FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeGEMMLowpFusedOffsetOutputUint8Dataset(),
- framework::dataset::make("DataType", { DataType::QASYMM8 })),
- framework::dataset::make("reshape_b_only_on_first_run", { true, false })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture, framework::DatasetMode::NIGHTLY,
+ combine(datasets::LargeGEMMLowpFusedOffsetOutputUint8Dataset(),
+ make("DataType", { DataType::QASYMM8 }),
+ make("reshape_b_only_on_first_run", { true, false })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
@@ -141,8 +146,9 @@
TEST_SUITE(QASYMM8_SIGNED)
using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInt8Fixture =
GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, false, int8_t, int8_t>;
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInt8Fixture, framework::DatasetMode::ALL, combine(datasets::SmallGEMMLowpFusedOffsetOutputInt8Dataset(),
- framework::dataset::make("DataType", { DataType::QASYMM8_SIGNED })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInt8Fixture, framework::DatasetMode::ALL,
+ combine(datasets::SmallGEMMLowpFusedOffsetOutputInt8Dataset(),
+ make("DataType", { DataType::QASYMM8_SIGNED })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
@@ -185,24 +191,24 @@
TEST_SUITE(QASYMM8)
-const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
- 3)
- * framework::dataset::make("min", 0) * framework::dataset::make("max", 255) * framework::dataset::make("addBias", { false, true });
+const auto quantize_down_int32_to_uint8_scale_cases = make("result_offset", -2, 1) * make("result_mult_int", 1, 2) * make("result_shift", 2, 3)
+ * make("min", 0) * make("max", 255) * make("addBias", { false, true });
-const auto quantize_down_int32_to_uint8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1,
- 2)
- * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 173) * framework::dataset::make("addBias", { false, true });
+const auto quantize_down_int32_to_uint8_scale_relu_cases = make("result_offset", -2, 1) * make("result_mult_int", 1, 2)
+ * make("result_shift", 2, 3) * make("min", 0, 2) * make("max", 171, 173) * make("addBias", { false, true });
using CLGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage>;
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL,
+ combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases))
{
// Validate output
validate(CLAccessor(_target), _reference);
}
TEST_SUITE(BoundedReLu)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL,
+ combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
{
// Validate output
validate(CLAccessor(_target), _reference);
@@ -213,24 +219,24 @@
TEST_SUITE(QASYMM8_SIGNED)
-const auto quantize_down_int32_to_int8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
- 3)
- * framework::dataset::make("min", -128) * framework::dataset::make("max", 127) * framework::dataset::make("addBias", { false, true });
+const auto quantize_down_int32_to_int8_scale_cases = make("result_offset", -2, 1) * make("result_mult_int", 1, 2) * make("result_shift", 2, 3)
+ * make("min", -128) * make("max", 127) * make("addBias", { false, true });
-const auto quantize_down_int32_to_int8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1,
- 2)
- * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", -100, -98) * framework::dataset::make("max", 71, 73) * framework::dataset::make("addBias", { false, true });
+const auto quantize_down_int32_to_int8_scale_relu_cases = make("result_offset", -2, 1) * make("result_mult_int", 1, 2)
+ * make("result_shift", 2, 3) * make("min", -100, -98) * make("max", 71, 73) * make("addBias", { false, true });
using CLGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage>;
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_cases))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL,
+ combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_cases))
{
// Validate output
validate(CLAccessor(_target), _reference);
}
TEST_SUITE(BoundedReLu)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_relu_cases))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL,
+ combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_relu_cases))
{
// Validate output
validate(CLAccessor(_target), _reference);
@@ -247,13 +253,14 @@
GEMMLowpQuantizeDownInt32ScaleByFloatValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage, uint8_t>;
FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMLowpQuantizeDownInt32ScaleByFloatFixture, framework::DatasetMode::ALL,
- combine(combine(combine(combine(combine(combine(framework::dataset::make("DataType", DataType::QASYMM8),
- datasets::TinyShapes()),
- framework::dataset::make("result_real_multiplier", 0.33f)),
- framework::dataset::make("result_offset", 2, 3)),
- framework::dataset::make("min", 0)),
- framework::dataset::make("max", 255)),
- framework::dataset::make("addBias", { false, true })))
+ combine(
+ make("DataType", DataType::QASYMM8),
+ datasets::TinyShapes(),
+ make("result_real_multiplier", 0.33f),
+ make("result_offset", 2, 3),
+ make("min", 0),
+ make("max", 255),
+ make("addBias", { false, true })))
{
// Validate output
validate(CLAccessor(_target), _reference);
@@ -264,13 +271,14 @@
using CLGEMMLowpQuantizeDownInt32ScaleByFloatFixture_Signed =
GEMMLowpQuantizeDownInt32ScaleByFloatValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage, int8_t>;
FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMLowpQuantizeDownInt32ScaleByFloatFixture_Signed, framework::DatasetMode::ALL,
- combine(combine(combine(combine(combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED),
- datasets::TinyShapes()),
- framework::dataset::make("result_real_multiplier", 0.33f)),
- framework::dataset::make("result_offset", 2, 3)),
- framework::dataset::make("min", -128)),
- framework::dataset::make("max", 127)),
- framework::dataset::make("addBias", { false, true })))
+ combine(
+ make("DataType", DataType::QASYMM8_SIGNED),
+ datasets::TinyShapes(),
+ make("result_real_multiplier", 0.33f),
+ make("result_offset", 2, 3),
+ make("min", -128),
+ make("max", 127),
+ make("addBias", { false, true })))
{
// Validate output
validate(CLAccessor(_target), _reference);