diff options
author | Vidhya Sudhan Loganathan <vidhyasudhan.loganathan@arm.com> | 2019-02-01 09:49:50 +0000 |
---|---|---|
committer | VidhyaSudhan Loganathan <vidhyasudhan.loganathan@arm.com> | 2019-02-01 16:26:32 +0000 |
commit | f8b652052960e19d2d2d858600780b594d108cc8 (patch) | |
tree | 2f59ced448d333aa82a25fe3c7f884e6d2e86679 | |
parent | aa3fc130acb84f7262fa47d9b5107f0afcf105e0 (diff) | |
download | ComputeLibrary-f8b652052960e19d2d2d858600780b594d108cc8.tar.gz |
COMPMID-1911 : NEON QASYMM8 add should be rounding to nearest, tie to even
Change-Id: I10f702be6d5555b6b89f3413d07b62b2a961a1c9
Signed-off-by: Vidhya Sudhan Loganathan <vidhyasudhan.loganathan@arm.com>
Reviewed-on: https://review.mlplatform.org/602
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
-rw-r--r-- | src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp | 19 | ||||
-rw-r--r-- | tests/validation/NEON/ArithmeticAddition.cpp | 6 |
2 files changed, 22 insertions, 3 deletions
diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp index bd52d4de2b..ffa578f40e 100644 --- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp +++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp @@ -166,7 +166,6 @@ void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0); const float output_scale = out->info()->quantization_info().scale; - const float invoutput_scale = 1.f / output_scale; const int output_offset = out->info()->quantization_info().offset; const float32x4_t vscale1 = vdupq_n_f32(in1->info()->quantization_info().scale); @@ -230,10 +229,17 @@ void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor const int32x4x4_t rf = { { +#ifdef __aarch64__ + vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)), + vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)), + vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)), + vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)), +#else //__aarch64__ vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)), vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)), vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)), vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)), +#endif //__aarch64__ } }; @@ -246,7 +252,7 @@ void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor for(; x < window_end_x; ++x) { const float afs = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale; - *(output_ptr + x) = std::max<int32_t>(0, std::min<int32_t>((afs + bfs) * invoutput_scale + output_offset, 255)); + *(output_ptr + x) = out->info()->quantization_info().quantize((afs + bfs),RoundingPolicy::TO_NEAREST_UP); } }, broadcast_input, non_broadcast_input, output); @@ -300,10 +306,17 @@ void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor const int32x4x4_t rf = { { +#ifdef __aarch64__ + vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)), + vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)), + vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)), + vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)), +#else //__aarch64__ vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)), vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)), vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)), vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)), +#endif //__aarch64__ } }; @@ -317,7 +330,7 @@ void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor { const float afs = static_cast<int32_t>((*(input1_ptr + x)) - input1_qinfo.offset) * input1_qinfo.scale; const float bfs = static_cast<int32_t>((*(input2_ptr + x)) - input2_qinfo.offset) * input2_qinfo.scale; - *(output_ptr + x) = std::max<int32_t>(0, std::min<int32_t>((afs + bfs) * invoutput_scale + output_offset, 255)); + *(output_ptr + x) = out->info()->quantization_info().quantize((afs + bfs),RoundingPolicy::TO_NEAREST_UP); } }, input1, input2, output); diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp index 0816283d83..bad0b33d37 100644 --- a/tests/validation/NEON/ArithmeticAddition.cpp +++ b/tests/validation/NEON/ArithmeticAddition.cpp @@ -43,7 +43,9 @@ namespace validation { namespace { +#ifndef __aarch64__ constexpr AbsoluteTolerance<float> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ +#endif //__aarch64__ /** Input data sets **/ const auto ArithmeticAdditionU8Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)), framework::dataset::make("DataType", @@ -278,7 +280,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) }))) { // Validate output +#ifdef __aarch64__ + validate(Accessor(_target), _reference); +#else //__aarch64__ validate(Accessor(_target), _reference, tolerance_qasymm8); +#endif //__aarch64__ } TEST_SUITE_END() // QASYMM8 TEST_SUITE_END() // Quantized |