diff options
| -rw-r--r-- | src/core/NEON/kernels/NEReductionOperationKernel.cpp | 11 | ||||
| -rw-r--r-- | tests/validation/NEON/ArgMinMax.cpp | 40 |
2 files changed, 27 insertions, 24 deletions
diff --git a/src/core/NEON/kernels/NEReductionOperationKernel.cpp b/src/core/NEON/kernels/NEReductionOperationKernel.cpp index 64e3cfe404..d88717a396 100644 --- a/src/core/NEON/kernels/NEReductionOperationKernel.cpp +++ b/src/core/NEON/kernels/NEReductionOperationKernel.cpp @@ -74,10 +74,13 @@ uint32x4x4_t calculate_index(uint32_t idx, uint8x16_t a, uint8x16_t b, uint32x4x { mask_u8 = wrapper::vclt(b, a); } - mask.val[0] = wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(mask_u8)))); - mask.val[1] = wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(mask_u8)))); - mask.val[2] = wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(mask_u8)))); - mask.val[3] = wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(mask_u8)))); + auto wide_u16_1 = wrapper::vorr(vshll_n_u8(wrapper::vgetlow(mask_u8), 8), wrapper::vmovl(wrapper::vgetlow(mask_u8))); + auto wide_u16_2 = wrapper::vorr(vshll_n_u8(wrapper::vgethigh(mask_u8), 8), wrapper::vmovl(wrapper::vgethigh(mask_u8))); + mask.val[0] = wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_1), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_1))); + mask.val[1] = wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_1), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_1))); + mask.val[2] = wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_2), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_2))); + mask.val[3] = wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_2), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_2))); + uint32x4x4_t vec_idx = { { { idx + 0, idx + 1, idx + 2, idx + 3 }, { idx + 4, idx + 5, idx + 6, idx + 7 }, { idx + 8, idx + 9, idx + 10, idx + 11 }, diff --git a/tests/validation/NEON/ArgMinMax.cpp b/tests/validation/NEON/ArgMinMax.cpp index 611495a41d..bf7dbfe6b7 100644 --- a/tests/validation/NEON/ArgMinMax.cpp +++ b/tests/validation/NEON/ArgMinMax.cpp @@ -113,53 +113,53 @@ FIXTURE_DATA_TEST_CASE(RunLarge, TEST_SUITE_END() // FP16 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template <typename T> -using NEArgMinMaxQuantizedValidationFixture = ArgMinMaxValidationQuantizedFixture<Tensor, Accessor, NEArgMinMaxLayer, T>; - -TEST_SUITE(QASYMM8) +TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, - NEArgMinMaxQuantizedValidationFixture<uint8_t>, + NEArgMinMaxValidationFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) + combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, - NEArgMinMaxQuantizedValidationFixture<uint8_t>, + NEArgMinMaxValidationFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) + combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) { // Validate output validate(Accessor(_target), _reference); } -TEST_SUITE_END() // QASYMM8 +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // Float -TEST_SUITE(FP32) +template <typename T> +using NEArgMinMaxQuantizedValidationFixture = ArgMinMaxValidationQuantizedFixture<Tensor, Accessor, NEArgMinMaxLayer, T>; + +TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, - NEArgMinMaxValidationFixture<float>, + NEArgMinMaxQuantizedValidationFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), + framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, - NEArgMinMaxValidationFixture<float>, + NEArgMinMaxQuantizedValidationFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), + framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) { // Validate output validate(Accessor(_target), _reference); } -TEST_SUITE_END() // FP32 -TEST_SUITE_END() // Float +TEST_SUITE_END() // QASYMM8 TEST_SUITE_END() // ArgMinMax TEST_SUITE_END() // NEON } // namespace validation |