diff options
Diffstat (limited to 'tests/validation/reference/PixelWiseMultiplication.cpp')
-rw-r--r-- | tests/validation/reference/PixelWiseMultiplication.cpp | 117 |
1 files changed, 116 insertions, 1 deletions
diff --git a/tests/validation/reference/PixelWiseMultiplication.cpp b/tests/validation/reference/PixelWiseMultiplication.cpp index 3e21fca72a..0450991f61 100644 --- a/tests/validation/reference/PixelWiseMultiplication.cpp +++ b/tests/validation/reference/PixelWiseMultiplication.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 ARM Limited. + * Copyright (c) 2017-2020 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,6 +43,8 @@ struct is_floating_point namespace { +constexpr float scale1_constant = 1.f; + /** Compute the result of `src1 * src2 * scale`. The result type always matches the type of @p src2. * * @param[in] src1 An input value. Data types supported: U8/S16/F16/F32. @@ -89,6 +91,90 @@ T3 mul(const T1 src1, const T2 src2, float scale, ConvertPolicy convert_policy, } } +template <> +int32_t mul(const int32_t src1, const int32_t src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) +{ + const int64_t intermediate_val = static_cast<int64_t>(src1) * static_cast<int64_t>(src2); + + if(std::abs(scale - scale1_constant) < 0.00001f) + { + // Use bit-accurate integer arithmetic for scale == 1 + // Apply conversion + if(convert_policy == ConvertPolicy::SATURATE) + { + return saturate_cast<int32_t>(intermediate_val); + } + else + { + // Correct wrapping behaviour for int32_t + const auto i32_hi = static_cast<int64_t>(std::numeric_limits<int32_t>::max()); + const auto i32_lo = static_cast<int64_t>(std::numeric_limits<int32_t>::lowest()); + const auto i32_wi = static_cast<int64_t>(1) << 32; + int64_t wrapped_rounded_val = intermediate_val - i32_wi * static_cast<int64_t>(support::cpp11::trunc(static_cast<double>(intermediate_val) / i32_wi)); + if(wrapped_rounded_val <= i32_hi) + { + return static_cast<int32_t>(wrapped_rounded_val); + } + else + { + // Values beyond i32_hi wrap around to negatives + return static_cast<int32_t>((wrapped_rounded_val - i32_hi) + i32_lo - 1); + } + } + } + else + { + // Use double arithmetic for scale != 1; may not be bit-accurate + // Apply scaling + // scale == 1 / 2^scale_exponent + int scale_exponent = 0; + std::frexp(scale, &scale_exponent); + // Store the positive exponent. We know that we compute 1/2^n + // Additionally we need to subtract 1 to compensate that frexp used a mantissa of 0.5 + scale_exponent = std::abs(scale_exponent - 1); + const double scale_inv = static_cast<int64_t>(1) << scale_exponent; + const double val = intermediate_val / scale_inv; + // Apply rounding + double rounded_val = 0; + switch(rounding_policy) + { + case(RoundingPolicy::TO_ZERO): + rounded_val = support::cpp11::trunc(val); + break; + case(RoundingPolicy::TO_NEAREST_UP): + rounded_val = round_half_up(val); + break; + case(RoundingPolicy::TO_NEAREST_EVEN): + rounded_val = round_half_even(val); + break; + default: + ARM_COMPUTE_ERROR("Unsupported rounding policy"); + } + // Apply conversion + if(convert_policy == ConvertPolicy::SATURATE) + { + return saturate_cast<int32_t>(rounded_val); + } + else + { + // Correct wrapping behaviour for int32_t + const auto i32_hi = static_cast<double>(std::numeric_limits<int32_t>::max()); + const auto i32_lo = static_cast<double>(std::numeric_limits<int32_t>::lowest()); + const auto i32_wi = static_cast<double>(static_cast<int64_t>(1) << 32); + double wrapped_rounded_val = rounded_val - i32_wi * std::floor(rounded_val / i32_wi); + if(wrapped_rounded_val <= i32_hi) + { + return static_cast<int32_t>(wrapped_rounded_val); + } + else + { + // Values beyond i32_hi wrap around to negatives + return static_cast<int32_t>((wrapped_rounded_val - i32_hi) + i32_lo - 1); + } + } + } +} + template <size_t dim> struct BroadcastUnroll { @@ -178,6 +264,34 @@ SimpleTensor<uint8_t> pixel_wise_multiplication(const SimpleTensor<uint8_t> &src } template <> +SimpleTensor<int16_t> pixel_wise_multiplication(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, + DataType dt_out, const QuantizationInfo &qout) +{ + SimpleTensor<int16_t> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dt_out, 1, qout); + + if(src1.data_type() == DataType::QASYMM8 && src2.data_type() == DataType::QASYMM8) + { + SimpleTensor<float> src1_tmp = convert_from_asymmetric(src1); + SimpleTensor<float> src2_tmp = convert_from_asymmetric(src2); + SimpleTensor<float> dst_tmp = pixel_wise_multiplication<float, float, float>(src1_tmp, src2_tmp, scale, convert_policy, rounding_policy, DataType::F32, qout); + dst = convert_to_symmetric<int16_t>(dst_tmp, qout); + } + else + { + if(scale < 0) + { + ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative"); + } + + Coordinates id_src1{}; + Coordinates id_src2{}; + Coordinates id_dst{}; + BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst); + } + return dst; +} + +template <> SimpleTensor<int8_t> pixel_wise_multiplication(const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout) { @@ -236,6 +350,7 @@ SimpleTensor<int16_t> pixel_wise_multiplication(const SimpleTensor<int16_t> &src // clang-format off template SimpleTensor<int16_t> pixel_wise_multiplication(const SimpleTensor<uint8_t> &src1, const SimpleTensor<int16_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout); template SimpleTensor<int32_t> pixel_wise_multiplication(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout); +template SimpleTensor<int32_t> pixel_wise_multiplication(const SimpleTensor<int32_t> &src1, const SimpleTensor<int32_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout); template SimpleTensor<float> pixel_wise_multiplication(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout); template SimpleTensor<half_float::half> pixel_wise_multiplication(const SimpleTensor<half_float::half> &src1, const SimpleTensor<half_float::half> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout); // clang-format on |