COMPMID-661: Add avgpool-uint8 support. Optimize avgpool-fp32 for Bifrost. (#13)

Change-Id: I32ba6afbac6694ffa053dd16f03a1b3d14627a19
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/94857
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 47 insertions, 12 deletions

diff --git a/tests/validation/CL/PoolingLayer.cpp b/tests/validation/CL/PoolingLayer.cpp
index 7038f2c34e..b3d56122db 100644
--- a/tests/validation/CL/PoolingLayer.cpp
+++ b/tests/validation/CL/PoolingLayer.cpp
@@ -43,19 +43,26 @@ namespace validation
 {
 namespace
 {
-/** Input data set for float data types */
+/** Input data set for floating-point data types */
 const auto PoolingLayerDatasetFP = combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { 2, 3, 4, 7, 9 })),
                                                    framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })),
                                            framework::dataset::make("ExcludePadding", { true, false }));
 
-/** Input data set for quantized data types */
+/** Input data set for fixed-point data types */
 const auto PoolingLayerDatasetQS = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { 2, 3 })),
                                                    framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })),
                                            framework::dataset::make("ExcludePadding", { true, false }));
-constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for float types */
-constexpr AbsoluteTolerance<float> tolerance_f16(0.01f);  /**< Tolerance value for comparing reference's output against implementation's output for float types */
-constexpr AbsoluteTolerance<float> tolerance_qs8(3);      /**< Tolerance value for comparing reference's output against implementation's output for quantized input */
-constexpr AbsoluteTolerance<float> tolerance_qs16(6);     /**< Tolerance value for comparing reference's output against implementation's output for quantized input */
+
+/** Input data set for asymmetric data type */
+const auto PoolingLayerDatasetQASYMM8 = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { 2, 3 })),
+                                                        framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })),
+                                                framework::dataset::make("ExcludePadding", { true, false }));
+
+constexpr AbsoluteTolerance<float>   tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for 32-bit floating-point type */
+constexpr AbsoluteTolerance<float>   tolerance_f16(0.01f);  /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */
+constexpr AbsoluteTolerance<float>   tolerance_qs16(6);     /**< Tolerance value for comparing reference's output against implementation's output for 16-bit fixed-point type */
+constexpr AbsoluteTolerance<float>   tolerance_qs8(3);      /**< Tolerance value for comparing reference's output against implementation's output for 8-bit fixed-point type */
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1);  /**< Tolerance value for comparing reference's output against implementation's output for 8-bit asymmetric type */
 } // namespace
 
 TEST_SUITE(CL)
@@ -64,12 +71,13 @@ TEST_SUITE(PoolingLayer)
 // *INDENT-OFF*
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
-               framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),
+               framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),     // Mismatching data type
                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),
-                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QS8, 4),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QS8, 4),     // Mismatching fixed point position
                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QS16, 11),
-                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),
-                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),     // Invalid pad/size combination
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),     // Invalid pad/size combination
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QASYMM8, 0), // Invalid parameters
                                                      }),
                framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F16, 0),
                                                        TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 0),
@@ -77,6 +85,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
                                                        TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::QS16, 11),
                                                        TensorInfo(TensorShape(30U, 11U, 2U), 1, DataType::F32, 0),
                                                        TensorInfo(TensorShape(25U, 16U, 2U), 1, DataType::F32, 0),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QASYMM8, 0),
                                                      })),
                framework::dataset::make("PoolInfo",  { PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 0, 0)),
                                                        PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 0, 0)),
@@ -84,8 +93,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
                                                        PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 0, 0)),
                                                        PoolingLayerInfo(PoolingType::AVG, 2, PadStrideInfo(1, 1, 2, 0)),
                                                        PoolingLayerInfo(PoolingType::AVG, 2, PadStrideInfo(1, 1, 0, 2)),
+                                                       PoolingLayerInfo(PoolingType::L2, 3, PadStrideInfo(1, 1, 0, 0)),
                                                       })),
-               framework::dataset::make("Expected", { true, false, true, false, true, true})),
+               framework::dataset::make("Expected", { true, false, true, false, true, true, true })),
                input_info, output_info, pool_info, expected)
 {
     ARM_COMPUTE_EXPECT(bool(CLPoolingLayer::validate(&input_info, &output_info, pool_info)) == expected, framework::LogLevel::ERRORS);
@@ -131,7 +141,7 @@ TEST_SUITE_END()
 template <typename T>
 using CLPoolingLayerFixedPointFixture = PoolingLayerValidationFixedPointFixture<CLTensor, CLAccessor, CLPoolingLayer, T>;
 
-TEST_SUITE(Quantized)
+TEST_SUITE(FixedPoint)
 TEST_SUITE(QS8)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS,
                                                                                                                        framework::dataset::make("DataType", DataType::QS8))),
@@ -167,6 +177,31 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLPoolingLayerFixedPointFixture<int16_t>, frame
 TEST_SUITE_END()
 TEST_SUITE_END()
 
+TEST_SUITE(Quantized)
+
+template <typename T>
+using CLPoolingLayerQuantizedFixture = PoolingLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLPoolingLayer, T>;
+
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQASYMM8,
+                                                                                                                       framework::dataset::make("DataType", DataType::QASYMM8))),
+                                                                                                               framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 127),
+                                                                                                                       QuantizationInfo(7.f / 255, 123)
+                                                                                                                                                            })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLPoolingLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), combine(PoolingLayerDatasetQASYMM8,
+                                                                                                                   framework::dataset::make("DataType", DataType::QASYMM8))),
+                                                                                                                   framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
 TEST_SUITE_END()
 TEST_SUITE_END()
 } // namespace validation