diff options
author | Viet-Hoa Do <viet-hoa.do@arm.com> | 2023-04-26 15:38:45 +0100 |
---|---|---|
committer | Viet-Hoa Do <viet-hoa.do@arm.com> | 2023-05-02 08:57:02 +0000 |
commit | a62129a02397ba87171ebf4477795f628dcec0f6 (patch) | |
tree | 91e53cc8982d9f16e66db53f81830fb05da83596 /tests/validation/fixtures | |
parent | f0ff76dbfc9137d0dfc5e99666e24c7a2ca8b072 (diff) | |
download | ComputeLibrary-a62129a02397ba87171ebf4477795f628dcec0f6.tar.gz |
Fix fully connected and matmul mismatches
* There is an issue with quantized fully connected and matmul
when the lower bound of bounded ReLU is negative.
* Use int32_t for the calculation of min/max quantized value
rather than PixelValue to avoid this issue.
Partially resolves: COMPMID-5996
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I7b22e9d56a2441fc6a4c5c4e627f57d6e00d6ff1
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9502
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'tests/validation/fixtures')
-rw-r--r-- | tests/validation/fixtures/MatMulFixture.h | 16 |
1 files changed, 6 insertions, 10 deletions
diff --git a/tests/validation/fixtures/MatMulFixture.h b/tests/validation/fixtures/MatMulFixture.h index f8f038af3f..15719024b1 100644 --- a/tests/validation/fixtures/MatMulFixture.h +++ b/tests/validation/fixtures/MatMulFixture.h @@ -27,6 +27,7 @@ #include "arm_compute/core/Types.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "src/core/utils/quantization/AsymmHelpers.h" #include "tests/framework/Fixture.h" #include "tests/validation/reference/ActivationLayer.h" #include "tests/validation/reference/GEMM.h" @@ -162,16 +163,16 @@ protected: template <typename TT> typename std::enable_if<!std::is_integral<TT>::value, SimpleTensor<TT>>::type - compute_reference_gemm(const SimpleTensor<TT> &a, const SimpleTensor<TT> &b, const SimpleTensor<TT> &c, float alpha, float beta, const ActivationLayerInfo &act_info, const QuantizationInfo &o_qinfo) + compute_reference_gemm(const SimpleTensor<TT> &a, const SimpleTensor<TT> &b, const SimpleTensor<TT> &c, float alpha, float beta, const QuantizationInfo &o_qinfo) { - ARM_COMPUTE_UNUSED(act_info, o_qinfo); + ARM_COMPUTE_UNUSED(o_qinfo); return reference::gemm(a, b, c, alpha, beta); } template <typename TT> typename std::enable_if<std::is_integral<TT>::value, SimpleTensor<TT>>::type - compute_reference_gemm(const SimpleTensor<TT> &a, const SimpleTensor<TT> &b, const SimpleTensor<TT> &c, float alpha, float beta, const ActivationLayerInfo &act_info, const QuantizationInfo &o_qinfo) + compute_reference_gemm(const SimpleTensor<TT> &a, const SimpleTensor<TT> &b, const SimpleTensor<TT> &c, float alpha, float beta, const QuantizationInfo &o_qinfo) { ARM_COMPUTE_UNUSED(alpha, beta); @@ -187,17 +188,12 @@ protected: std::vector<int32_t> output_multipliers{ output_multiplier }; std::vector<int32_t> output_shifts{ output_shift }; - PixelValue output_min{}; - PixelValue output_max{}; - std::tie(output_min, output_max) = quantization::get_quantized_asymmetric_output_min_max( - o_qinfo, act_info, a.data_type()); - const auto tmp = reference::gemmlowp_matrix_multiply_core<int32_t>( a, b, c.shape(), aq.offset, bq.offset); auto output = reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, TT>( tmp, output_multipliers, output_shifts, oq.offset, - output_min.get<int32_t>(), output_max.get<int32_t>()); + std::numeric_limits<int32_t>::lowest(), std::numeric_limits<int32_t>::max()); output.quantization_info(o_qinfo); return output; @@ -253,7 +249,7 @@ protected: // Setting beta to 0 will effectively disable C for the // computation of the reference: alpha * A * B + 0 * C // Use transposed tensors if boolean enabled else use original tensors - auto result = compute_reference_gemm<T>((transpose_a) ? a_transposed : a, (transpose_b) ? b_transposed : b, c, 1.0f, 0.f, act_info, o_qinfo); + auto result = compute_reference_gemm<T>((transpose_a) ? a_transposed : a, (transpose_b) ? b_transposed : b, c, 1.0f, 0.f, o_qinfo); result = reference::activation_layer<T>(result, act_info, o_qinfo); |